VLA Simulator — 1-Click VLA Simulation on AWS

Run Vision-Language-Action (VLA) robot simulation workloads on AWS GPU instances with a single command. Supports NVIDIA GR00T N1.7, GR00T N1.6 (GR1 humanoid and Unitree G1 whole-body loco-manipulation), π0.5 (openpi), OpenVLA-OFT, LAP-3B, MolmoAct2 (Ai2), and RLDX-1 (RLWRLD) — plus an OpenArm Lift-Cube target in Isaac Lab for scripted ACT demo collection.

See Showcase — VLA Rollouts for sample rollouts across the trained policies and the OpenArm scripted collection run, spanning a range of LIBERO / RoboCasa verbs.

Overview

Feature	Detail
Models	GR00T N1.7-LIBERO, GR00T N1.6-3B (GR1), GR00T N1.6-G1 (Unitree G1 loco-manip), π0.5 (pi05_libero), OpenVLA-OFT (LIBERO-10), LAP-3B (LIBERO-Spatial), MolmoAct2 (Ai2 Action Reasoning Model, LIBERO-Goal), RLDX-1 (RLWRLD — LIBERO / SimplerEnv Fractal / RoboCasa GR-1 Tabletop)
Simulation	LIBERO / RoboCasa (robosuite + MuJoCo, headless EGL); Isaac Lab (OpenArm Lift-Cube collection)
Deploy	AWS CDK + EC2 GPU (g6/g5, us-east-1; OpenArm needs a 4-GPU `.12xl`)
Results	S3 (MP4 video + summary) + SNS email
Cleanup	Auto-terminate EC2; run `vlasim.py destroy` for stack teardown

Supported VLA Combinations

`--vla`	`--libero-suite`	Model	Sim Environment	Robot	Stack
`gr00t`	—	GR00T N1.7-LIBERO	LIBERO-10 kitchen tasks	Franka Panda (7-DOF)	`GR00T-Demo`
`gr00t-gr1`	—	GR00T N1.6-3B	RoboCasa GR1 tabletop tasks	Fourier GR1 humanoid (22-DOF)	`GR00T-GR1-Demo`
`gr00t-g1`	—	GR00T N1.6-G1 (community re-finetune)	GR00T-WholeBodyControl loco-manip (GEAR WBC + MuJoCo)	Unitree G1 humanoid (whole-body)	`GR00T-G1-Demo`
`pi`	—	π0.5 (pi05_libero)	LIBERO spatial/object	Franka Panda (7-DOF)	`Pi-Demo`
`openvla-oft`	`spatial`	OpenVLA-OFT-7B (LIBERO-Spatial fine-tune)	LIBERO-Spatial	Franka Panda (7-DOF)	`OpenVLA-OFT-Spatial-Demo`
`openvla-oft`	`object`	OpenVLA-OFT-7B (LIBERO-Object fine-tune)	LIBERO-Object	Franka Panda (7-DOF)	`OpenVLA-OFT-Object-Demo`
`openvla-oft`	`goal`	OpenVLA-OFT-7B (LIBERO-Goal fine-tune)	LIBERO-Goal	Franka Panda (7-DOF)	`OpenVLA-OFT-Goal-Demo`
`openvla-oft`	`10` (default, alias `long`)	OpenVLA-OFT-7B (LIBERO-10 fine-tune)	LIBERO-10 long-horizon	Franka Panda (7-DOF)	`OpenVLA-OFT-Demo`
`lap`	—	LAP-3B (PaliGemma-3B + Flow Matching, JAX)	LIBERO-Spatial	Franka Panda (7-DOF)	`LAP-Demo`
`molmoact2`	—	MolmoAct2 (Ai2 Action Reasoning Model, 5B; Molmo2-ER VLM + flow-matching action expert)⁵	LIBERO-Goal	Franka Panda (7-DOF)	`MolmoAct2-Demo`
`rldx`	—	RLDX-1 (RLWRLD MSAT / Qwen3-VL-8B, eager)¹	LIBERO-10 long-horizon	Franka Panda (7-DOF)	`RLDX-Demo`
`rldx-simpler`	—	RLDX-1 (RLWRLD MSAT / Qwen3-VL-8B, eager)¹ ²	SimplerEnv Google-VM (Fractal)	Google Robot (OXE_FRACTAL real-robot embodiment)	`RLDX-Simpler-Demo`
`rldx-gr1`	—	RLDX-1 (RLWRLD MSAT / Qwen3-VL-8B, eager)¹ ³	RoboCasa GR-1 Tabletop (24-task)	GR-1 humanoid (bimanual + waist)	`RLDX-GR1-Demo`
`rldx-kitchen`	—	RLDX-1 (RLWRLD MSAT / Qwen3-VL-8B, eager)¹ ⁴	RoboCasa Kitchen (24-task)	PandaOmron (single-arm mobile manipulator)	`RLDX-Kitchen-Demo`
`openarm-lift-act`	—	Scripted state machine (ACT data collection)	Isaac Lab Lift-Cube	OpenArm (unimanual, 7-DOF + gripper)	`OpenArm-Lift-ACT-Demo`

¹ RLDX-1 weights are under the RLWRLD Model License v1.0 (non-commercial); "simulation benchmarking" is an explicit intended use. AWS-internal enablement / benchmarking only — not for customer-facing commercial positioning. Weights are downloaded from Hugging Face at runtime (not vendored).

² rldx-simpler is the same RLDX-1 model on SimplerEnv (Google Fractal real-robot embodiment, OXE_FRACTAL) instead of LIBERO — the one target here that exercises a real-robot OXE embodiment rather than a simulated Panda/humanoid. Setup clones SimplerEnv + ManiSkill2_real2sim at runtime (MIT). End-to-end validated 2026-06-24 (g6.2xlarge L4, full deploy.py 1-shot smoke); checkpoint pinned at @f59c79e1. See the showcase and Expected Results.

³ rldx-gr1 is the same RLDX-1 model family on RoboCasa GR-1 Tabletop (RLWRLD's own RLDX-1-FT-GR1 checkpoint) — the only target here on a bimanual humanoid with a waist DOF (GR-1, GR1ArmsAndWaistFourierHands), distinct from all single-arm LIBERO/SimplerEnv clips. Renders via MuJoCo/robosuite (no Vulkan); the RoboCasa GR-1 tabletop-tasks suite is cloned at runtime (MIT). End-to-end validated 2026-06-25 (g5.2xlarge A10G 24GB, full deploy.py 1-shot smoke); checkpoint pinned at @d0277c5c. As source-verified, GR-1 needed no SimplerEnv-style code patch (dispatch + obs/action routing already correct at the pin) — only the shared dependency-pin fixes. Measured 0.80 (PnPCupToDrawerClose) and 0.20 (PnPMilkToMicrowaveClose) at n=5; reported paper SR = 58.7% (24-task average). See Expected Results.

⁴ rldx-kitchen is the same RLDX-1 model family on the RoboCasa Kitchen 24-task benchmark (RLWRLD's own RLDX-1-FT-ROBOCASA checkpoint) — a single-arm mobile manipulator (PandaOmron: a Franka Panda arm on an Omron LD mobile base) doing pick-and-place, doors, drawers and appliances across full kitchen scenes, distinct from the GR-1 humanoid and the LIBERO/SimplerEnv clips. Renders via MuJoCo/robosuite (no Vulkan); the RoboCasa kitchen suite (squarefk/robocasa, MIT) is cloned at runtime and its ~5 GB of scene assets download during setup. Source-verified clean like GR-1 (no dispatch/obs/action patch) with three RoboCasa-Kitchen-specific fixes found by reading source: (a) the cloned download_kitchen_assets.py has no -y argparse and prompts via input() → made non-interactive so it runs headless; (b) a dependency-pin reconciliation (the kitchen fork pins numpy 1.23.3/numba 0.56.4, but the import-reachable albumentations needs numpy>=1.24.4, so the GR-1-proven numpy 1.26.4/numba 0.61.2 stack is protected); (c) the env emits each of the 3 cameras at two resolutions (256 + 512) and the video recorder stacks all of them into a six-panel strip — narrowed to the 256-px policy views so the saved video is a clean three-panel clip (cosmetic; obs/policy/SR untouched). Checkpoint pinned at @e8279a38. End-to-end validated 2026-06-26 (g6.2xlarge L4 24 GB, full deploy.py 1-shot smoke); fixes (a)+(b) confirmed live (the ~5 GB asset download ran headless and the numpy/numba/albumentations stack resolved); fix (c) observed in the Gate-3 output and applied to the committed clips (re-validates live on next deploy). Two articulated tasks at n=5: OpenSingleDoor success_rate = 0.80 (4/5) and OpenDrawer success_rate = 0.80 (4/5); reported paper SR = 70.6% (24-task average). See Expected Results.

⁵ molmoact2 is Apache-2.0 (weights + code, commercial use OK). Unlike the LeRobot-native targets, it is not on upstream LeRobot — it is served from the allenai/lerobot fork (molmoact2-policy branch, pinned @a4f15bf3) installed via uv sync --locked (Python 3.12), running the public allenai/MolmoAct2-LIBERO checkpoint in-process on the sim GPU (no separate policy server). The official LIBERO eval recipe is float32 (use_amp=false) + norm_tag=libero, continuous action mode — so the ~20 GB fp32 checkpoint forces a ≥40 GB single GPU and ≥64 GiB host RAM (g6e.2xlarge+ L40S; g6e.xlarge's 32 GiB host RAM thrashes the checkpoint load and is excluded). End-to-end validated 2026-06-28 (g6e.2xlarge, us-west-2; full deploy.py 1-shot). libero_goal = 1.000 (2 tasks × 5 episodes); reported paper LIBERO average = 97.2%. See the showcase and Expected Results.

Showcase — VLA Rollouts

Sample rollouts captured directly from deploy.py runs and synced from each stack's S3 results bucket. GIFs are downsampled previews (288 px, 10 fps, ≤8 s); click-through to the full-quality MP4 in docs/showcase/ for the original frame rate and resolution.

`pi` — π0.5 on LIBERO (Franka Panda 7-DOF)

libero_spatial = 0.99 (10 tasks × 10 episodes), libero_object = 0.96 (10 tasks × 5 episodes) — g5.xlarge, validated 2026-04-22.

LIBERO-Spatial — pick black bowl between plate and ramekin	LIBERO-Object — pick milk and place in basket

MP4: `pi/libero-spatial-between-plate-ramekin-success.mp4`	MP4: `pi/libero-object-milk-success.mp4`

`openvla-oft` — OpenVLA-OFT on LIBERO-10 long-horizon

Each task is a two-stage instruction. Long-horizon means the policy must complete one sub-goal, recognise it, then proceed to the next. Verb diversity below shows the same checkpoint following structurally different instructions. g6.xlarge, validated 2026-05-04.

`put both the alphabet soup and the tomato sauce in the basket`	`turn on the stove and put the moka pot on it`

MP4: `openvla-oft/libero-10-soup-and-sauce-success.mp4`	MP4: `openvla-oft/libero-10-stove-moka-pot-success.mp4`

`put the white mug on the left plate and the yellow mug on the right`	`pick up the book and place it in the back compartment`

MP4: `openvla-oft/libero-10-mug-left-yellow-right-success.mp4`	MP4: `openvla-oft/libero-10-book-compartment-success.mp4`

`lap` — LAP-3B on LIBERO-Spatial

libero_spatial = 0.98 (10 tasks × 5 trials, g6.xlarge, validated 2026-05-17). Same scene as pi above, different policy — useful for side-by-side comparison. The cookie_box task is the one consistent failure mode (paper Table III range 85–95% leaves 1–2 tasks expected to miss).

Success — `pick up the black bowl between the plate and the ramekin`	Failure — `pick up the black bowl on the cookie box`

MP4: `lap/libero-spatial-between-plate-ramekin-success.mp4`	MP4: `lap/libero-spatial-cookie-box-failure.mp4`

`molmoact2` — MolmoAct2 (Ai2) on LIBERO-Goal (Franka Panda 7-DOF)

MolmoAct2 is Ai2's open Action Reasoning Model (5B; a Molmo2-ER VLM backbone + a flow-matching continuous-action expert), Apache-2.0 weights and code. Unlike the LeRobot-native models above, it is served from the allenai/lerobot fork (molmoact2-policy branch, not yet merged upstream), pinned at @a4f15bf3; the public allenai/MolmoAct2-LIBERO checkpoint is run in-process on the sim GPU. libero_goal = 1.000 (2 tasks × 5 episodes = 10/10), g6e.2xlarge (L40S 48 GB), float32 + norm_tag=libero, continuous action mode, validated 2026-06-28. Both clips below are the same put the bowl on the stove task from different LIBERO initial states (the only task rendered locally this run); the consistent place-on-burner across init states is what 1.000 looks like here.

Success — `put the bowl on the stove` (init state A)	Success — `put the bowl on the stove` (init state B)

MP4: `molmoact2/libero-goal-bowl-on-stove-success.mp4`	MP4: `molmoact2/libero-goal-bowl-on-stove-2-success.mp4`

`rldx` — RLDX-1 (RLWRLD) on LIBERO-10 long-horizon (Franka Panda 7-DOF)

RLDX-1 is a Korea-origin SOTA open VLA (MSAT action head on a Qwen3-VL-8B backbone; 97.8% LIBERO average in the paper). Served eager via a ZeroMQ policy server + sim client (two venvs), g6.xlarge L4 sm_89, validated 2026-06-20. Two LIBERO-10 long-horizon tasks (distinct scenes from the pi / lap spatial clips above): STUDY_SCENE1 book→caddy success_rate = 1.0 (5/5); KITCHEN_SCENE6 mug→microwave+close (articulated, two-stage) success_rate = 0.8 (4/5). Weights are non-commercial (RLWRLD Model License v1.0) — sim benchmarking only.

Success — `STUDY_SCENE1: pick up the book and place it in the back compartment of the caddy` (SR 1.0, 5/5)	Success — `KITCHEN_SCENE6: put the yellow and white mug in the microwave and close it` (SR 0.8, 4/5)

MP4: `rldx/libero10-study-book-caddy-success.mp4`	MP4: `rldx/libero10-kitchen-mug-microwave-success.mp4`
Tight, consistent: all 5 episodes complete in 20–23 env-steps.	Failure (1/5) — same task, hits the 90-step cap without closing:

	MP4: `rldx/libero10-kitchen-mug-microwave-failure.mp4` (clip is 2× speed)

Reading the success rate — "success" is a binary goal flag, not a quality score. LIBERO marks an episode success the moment its BDDL goal predicate is satisfied; it says nothing about how cleanly or how close to the failure boundary the policy got there. That distinction matters on long-horizon tasks:

STUDY_SCENE1 (SR 1.0) is genuinely robust — every episode finishes in 20–23 env-steps with margin to spare.

KITCHEN_SCENE6 (SR 0.8) is marginal even when it succeeds: successful episodes range from a clean 32 steps to a labored 47 steps, while the failure runs the full 90-step cap (a timeout, not a hard error). The scene also contains a second (grey) mug as a distractor next to the yellow/white target. A near-cap success is one perturbation away from a timeout — so treat 0.8 here as "works, but not yet deployment-margin," not as a 4-in-5 guarantee of reliable execution.

Takeaway for demos: pair the SR number with the step-count spread (in simulation_results.csv) and watch the failure clip — a high SR with episodes clustered near the step cap is a weaker result than the same SR with short, consistent runs.

`rldx-simpler` — RLDX-1 (RLWRLD) on SimplerEnv Google Fractal (real-robot embodiment)

Same RLDX-1 checkpoint family as rldx above, but evaluated on SimplerEnv (ManiSkill2_real2sim + SAPIEN) with the Google Robot OXE_FRACTAL embodiment — the one target here that exercises a real-robot Open X-Embodiment embodiment rather than a simulated Panda/humanoid. Served eager via the same ZeroMQ policy server + sim client (two venvs), g6.2xlarge L4 sm_89, validated 2026-06-24 via a full deploy.py 1-shot smoke. Two Google-VM tasks: google_robot_pick_coke_can success_rate = 1.0 (5/5) and google_robot_move_near success_rate = 1.0 (5/5). Weights are non-commercial (RLWRLD Model License v1.0) — sim benchmarking only.

Clips are shown at 0.25× (slowed 4×), with the playback rate overlaid bottom-left. These are atomic Google-robot tasks — each episode runs only 10–25 sim steps, so at native speed it finishes in ~1 second. The 4× slowdown makes the motion legible and paces these clips with the long-horizon GR-1 clips below (up to 720 steps). No frames are dropped or trimmed; only the timestamps are stretched.

Success — `google_robot_pick_coke_can` (SR 1.0, 5/5)	Success — `google_robot_move_near` (SR 1.0, 5/5)

MP4: `rldx-simpler/simpler-google-pick-coke-can-success.mp4`	MP4: `rldx-simpler/simpler-google-move-near-success.mp4`
Tight and consistent: episodes complete in 10–25 env-steps (shown at 0.25×).	Real-robot Fractal embodiment — distinct from every LIBERO/RoboCasa clip above (shown at 0.25×).

`rldx-gr1` — RLDX-1 (RLWRLD) on RoboCasa GR-1 Tabletop (bimanual humanoid + waist)

Same RLDX-1 family as rldx/rldx-simpler, but on RoboCasa GR-1 Tabletop with RLWRLD's own RLDX-1-FT-GR1 checkpoint — the only target here on a bimanual humanoid with a waist DOF (GR1ArmsAndWaistFourierHands: two 7-DOF arms + Fourier dexterous hands + waist), distinct from every single-arm LIBERO/SimplerEnv clip. Each clip is dual-view (left ego / right third-person); the policy controls both hands and the waist to manipulate articulated scene objects (a drawer, a microwave door). Served eager via the same ZeroMQ policy server + sim client (two venvs, MuJoCo/robosuite headless EGL — no Vulkan), g5.2xlarge A10G 24 GB, validated 2026-06-25 via a full deploy.py 1-shot smoke. Two articulated tasks at n=5: PnPCupToDrawerClose success_rate = 0.80 (4/5) and PnPMilkToMicrowaveClose success_rate = 0.20 (1/5) — the combined 5/10 is consistent with the paper's 58.7% 24-task average. Weights are non-commercial (RLWRLD Model License v1.0) — sim benchmarking only.

Success — `PnPCupToDrawerClose` (SR 0.80, 4/5)	Success — `PnPMilkToMicrowaveClose` (SR 0.20, 1/5)

MP4: `rldx-gr1/gr1-cup-drawer-success.mp4`	MP4: `rldx-gr1/gr1-milk-microwave-success.mp4`
Clean bimanual pick-place-close in 19 env-steps; the harder of the two tasks the policy reliably solves.	The rare success on the hard task (24 steps) — most runs time out (see failure clip).

Failure — `PnPMilkToMicrowaveClose` (4/5 of this task, step-cap timeout)

MP4: `rldx-gr1/gr1-milk-microwave-failure.mp4` (clip is 2× speed)
The dominant outcome on the microwave task: the run hits the 45-step cap (a timeout, not a hard error) without satisfying the goal predicate. SR 0.20 here means "rarely solved," not "1-in-5 reliable" — pair the number with the per-episode step counts in `simulation_results.csv`.

`rldx-kitchen` — RLDX-1 (RLWRLD) on RoboCasa Kitchen (PandaOmron mobile manipulator)

Same RLDX-1 family again, but on the RoboCasa Kitchen 24-task benchmark with RLWRLD's own RLDX-1-FT-ROBOCASA checkpoint — a single-arm mobile manipulator (PandaOmron: a Franka Panda arm on an Omron LD mobile base) operating articulated kitchen fixtures (cabinet doors, drawers) inside full procedurally-generated kitchen scenes, distinct from the GR-1 humanoid and every LIBERO/SimplerEnv clip. Each clip is a three-camera strip — the two external views (left / right) plus the wrist camera, exactly the views the policy consumes. Served eager via the same ZeroMQ policy server + sim client (two venvs, MuJoCo/robosuite headless EGL — no Vulkan), g6.2xlarge L4 24 GB, validated 2026-06-26 via a full deploy.py 1-shot smoke. Two articulated tasks at n=5: OpenSingleDoor success_rate = 0.80 (4/5) and OpenDrawer success_rate = 0.80 (4/5) — consistent with the paper's 70.6% 24-task average. Weights are non-commercial (RLWRLD Model License v1.0) — sim benchmarking only.

RoboCasa Kitchen's env emits each of the three cameras at two resolutions (a 256-px policy input + a raw 512-px copy), and the stock video recorder stacks all of them — so the unfiltered MP4 is a six-panel strip showing every camera twice. A target-gated runtime fix (FIX 10) narrows the recorder to the 256-px views, giving the clean three-panel video below. The clips here are post-processed to that same three-panel form; FIX 10 makes it the default on the next deploy.

Success — `OpenSingleDoor` (SR 0.80, 4/5)	Success — `OpenDrawer` (SR 0.80, 4/5)

MP4: `rldx-kitchen/kitchen-open-single-door-success.mp4`	MP4: `rldx-kitchen/kitchen-open-drawer-success.mp4`
Three-camera strip (left / right / wrist); the arm reaches the cabinet handle and swings the door open in 18 env-steps.	Distinct mobile-manipulator embodiment — the PandaOmron base + arm pulls the drawer open in 16 env-steps.

Failure — `OpenSingleDoor` (1/5 of this task, step-cap timeout)

MP4: `rldx-kitchen/kitchen-open-single-door-failure.mp4` (clip is 2× speed)
The one miss on the door task: the run hits the 45-step cap (a timeout, not a hard error) without latching the handle. SR 0.80 here means "usually solved" — pair the number with the per-episode step counts in `simulation_results.csv`.

`gr00t-gr1` — GR00T N1.6 on RoboCasa GR1 humanoid (22-DOF)

The GR1 humanoid is a different embodiment from Franka Panda — two arms, waist, Fourier dexterous hands. PosttrainPnPNovelFromPlateToBowlSplitA is in distribution for the post-trained N1.6 (~80% success), while PnPCanToDrawerClose is not supported by the pre-trained checkpoint and consistently fails — included to make the embodiment-coverage limitation visible. g6.12xlarge, validated 2026-04-10.

Success — `PosttrainPnPNovelFromPlateToBowlSplitA` (in-distribution)	Failure — `PnPCanToDrawerClose` (not supported by N1.6)

MP4: `gr00t-gr1/posttrain-pnp-plate-to-bowl-success.mp4`	MP4: `gr00t-gr1/pnp-can-to-drawer-failure.mp4`

`gr00t-g1` — GR00T N1.6 on Unitree G1 (whole-body loco-manipulation)

This is whole-body loco-manipulation, not tabletop: the dual-view clips show the G1 humanoid grasp the apple, then walk left on its own legs (GR00T-WholeBodyControl Balance/Walk ONNX policies) to a second table and place it on the plate — task pick up the apple, walk left and place the apple on the plate (LMPnPAppleToPlateDC_G1_gear_wbc). The left pane is the ego/close view, the right is third-person.

Checkpoint note. NVIDIA published no N1.7 Unitree G1 checkpoint, and the only public N1.6 G1 checkpoint (nvidia/GR00T-N1.6-G1-PnPAppleToPlate) does not reproduce its stated success rate under the released evaluation command — it scores 0/10, matching multiple open Isaac-GR00T issues with no upstream fix. The only public artifact that reproduces the task is a community re-finetune, cloudwalk-research/GR00T-N1.6-G1-PnPAppleToPlate (same GR00T N1.6 architecture, drop-in compatible), which scores success_rate = 0.3 (3/10) here. g6.12xlarge, n_envs=1, validated 2026-06-15.

Success — apple picked, walked left, placed on plate	Failure — grasp + walk OK, placement misses

MP4: `gr00t-g1/loco-manip-apple-to-plate-success.mp4`	MP4: `gr00t-g1/loco-manip-apple-to-plate-failure.mp4`

Optional: run a fine-tuned GR00T N1.7 Unitree-G1 adapter on the same sim

The same gr00t-g1 target can serve a fine-tuned GR00T N1.7 Unitree-G1 adapter instead of the N1.6 demo checkpoint. NVIDIA published no N1.7 G1 checkpoint (UNITREE_G1 is a post-train tag — the base gives ~0%), so this is a fine-tuned adapter. The default in models/gr00t-g1.yaml points server.hf_repo at a gated published checkpoint, andrewc76/GR00T-N1.7-G1-AppleToPlate (request access, then grant the deploy's Hugging Face token). To use your own instead, set server.hf_repo to your HF repo, or server.s3_ckpt_uri to an S3 bucket — collect demos with --collect (the N1.6 teacher above), fine-tune a G1 adapter on GR00T N1.7, and point the server at your merged checkpoint.

The interesting part is that the N1.7 server (Isaac-GR00T 65cc4a, Gr00tN1d7 / Cosmos-Reason2-2B) and the WBC sim env (Isaac-GR00T 77866395, the only commit where GR00T-WholeBodyControl still exists) cannot share one venv — so the deploy runs a 2-venv split across the ZMQ :8000 boundary and overlays the N1.7 wire codec (msgpack_numpy) onto the sim venv to keep obs/action serialization aligned. This is automatic when the server: block is present. See the comments in models/gr00t-g1.yaml for the design and the action-horizon consistency gotcha (a model/processor horizon mismatch is silent at export but crashes the rollout at step 0).

In our run, a G1 adapter fine-tuned on demos collected from the N1.6 teacher reached success_rate = 0.7 (7/10) on the same LMPnPAppleToPlateDC_G1_gear_wbc task — the N1.7 adapter clips below were produced exactly this way (your numbers will depend on your fine-tune).

N1.7 success — apple picked, walked left, placed on plate	N1.7 failure — grasp + walk OK, placement misses

MP4: `gr00t-g1/n17-loco-manip-apple-to-plate-success.mp4`	MP4: `gr00t-g1/n17-loco-manip-apple-to-plate-failure.mp4`

`gr00t` — GR00T N1.7 on LIBERO-10 (Franka Panda)

Video capture pending — KITCHEN_SCENE3/4 success rate is 1.0 on the validated runs (see Expected Results), but those rollout videos were not retained locally. Will backfill on the next --vla gr00t deploy.

`openarm-lift-act` — OpenArm unimanual Lift-Cube (Isaac Lab)

Not a learned-policy rollout. Unlike the entries above (each a trained VLA evaluated in LIBERO/RoboCasa), this target runs a scripted state machine (REST → APPROACH → GRASP → LIFT) on the OpenArm arm in Isaac Lab to collect successful pick-and-lift demonstrations — the camera + state + action HDF5 that an ACT policy is then trained on. The clip shows the data-collection rollout, not an evaluated policy. Training ACT on this dataset and showing a learned rollout is the next step.

Rendered at 768×768, captured during a --vla openarm-lift-act collection run (g6.12xlarge, 4× L4, eu-central-1, 2026-06-11). The third-person (table) view is the beauty shot; the wrist view is the gripper POV.

Third-person (table camera) — pick + lift cube	Wrist camera (gripper POV)

MP4: `openarm-lift-act/lift-cube-table-success.mp4`	MP4: `openarm-lift-act/lift-cube-wrist-success.mp4`

Architecture

deploy.py
  │
  ├─ generate.py          → assets/userdata/{vla}.sh
  │
  └─ CDK deploy
       ├─ VPC + SG
       ├─ S3 ResultsBucket (RETAIN)
       ├─ SNS + EmailSubscription
       ├─ IAM Role
       ├─ AzSelector Lambda  →  EC2 (g6.12xlarge / g5.xlarge)
       └─ WaitCondition      ←  cfn_signal from UserData

Two deployment modes:

Local mode — model runs directly on the EC2 instance (default)
Bridge mode — EC2 runs simulation env; model inference forwarded to an existing VLA Hub ECS endpoint

Quick Start

A single CLI (vlasim.py) wraps the whole setup into three commands. It runs on the standard library alone, so doctor works on a fresh host before any dependencies are installed.

# 1. See what's missing (read-only — safe to run anytime)
python vlasim.py doctor

# 2. One-time setup: installs Python + CDK deps and fills simulator-config.yaml
python vlasim.py init --email you@example.com
#   add --bootstrap to also run `cdk bootstrap` for your account/region

# 3. Deploy (runs doctor as a preflight first, then deploy.py)
python vlasim.py deploy --vla gr00t-g1

doctor checks: Python/Node/CDK versions and installs, AWS credentials, region support, CDK bootstrap, GPU On-Demand vCPU quota (per target), and notify_email. It exits non-zero if any check FAILs, so the deploy preflight stops a missing quota or un-bootstrapped account in seconds rather than mid-deploy (idle GPU billing).

$ python vlasim.py doctor --vla gr00t-g1
  [ OK ] Python 3.12.1 (>= 3.10)
  [ OK ] Node v22.22.1
  [ OK ] AWS credentials valid (account ..., region us-east-1)
  [ OK ] CDK bootstrap present in us-east-1
  [ OK ] On-Demand G/VR vCPU quota = 768 in us-east-1 (>= 4 for g6.xlarge ...)
  [FAIL] notify_email is still the placeholder
         -> run: python vlasim.py init --email YOU@EXAMPLE.COM

deploy and destroy forward every flag to deploy.py / destroy.py unchanged (--bridge, --collect, --libero-suite, --region, …). Use --skip-doctor to bypass the deploy preflight. Run python vlasim.py deploy --help for the full list.

Supported regions: us-east-1, us-west-2, ap-northeast-1, ap-northeast-2, eu-central-1

Manual setup (equivalent to init, if you prefer to run the steps yourself):

pip install -r requirements.txt          # boto3, pyyaml, jinja2
cd cdk && npm install && cd ..           # CDK deps (aws-cdk is pinned in cdk/package.json)
aws configure                            # or: export AWS_PROFILE=...
# then edit deployment.region / notify_email in simulator-config.yaml

Deploy Targets

Each target is one vlasim deploy --vla <target> (which runs the doctor preflight, then forwards every flag to deploy.py unchanged). The --email is only needed the first time, or if you skipped init.

# GR00T N1.7 — LIBERO kitchen tasks (~120 min)
python vlasim.py deploy --vla gr00t --email you@example.com

# GR00T N1.6 + GR1 humanoid — RoboCasa tabletop tasks (~90 min)
python vlasim.py deploy --vla gr00t-gr1 --email you@example.com

# GR00T N1.6 + Unitree G1 — whole-body loco-manipulation (GR00T-WholeBodyControl, ~20-30 min)
python vlasim.py deploy --vla gr00t-g1 --email you@example.com

# π0.5 — LIBERO spatial + object tasks (~4 hrs)
python vlasim.py deploy --vla pi --email you@example.com

# OpenVLA-OFT — LIBERO-10 long-horizon (~3 hrs, local mode only; default suite)
python vlasim.py deploy --vla openvla-oft --email you@example.com

# OpenVLA-OFT — LIBERO-Spatial short-horizon (~1.5 hrs)
python vlasim.py deploy --vla openvla-oft --libero-suite spatial --email you@example.com

# Other short-horizon suites: object, goal
python vlasim.py deploy --vla openvla-oft --libero-suite object --email you@example.com
python vlasim.py deploy --vla openvla-oft --libero-suite goal   --email you@example.com

# LAP-3B — LIBERO-Spatial (zero-shot cross-embodiment VLA, JAX policy server, ~1.5-2.5 hrs)
python vlasim.py deploy --vla lap --email you@example.com

# MolmoAct2 (Ai2) — LIBERO-Goal (Apache-2.0 5B Action Reasoning Model, fp32 in-process, g6e GPU ≥48GB, ~40-60 min)
# Use --region us-west-2 (g6e capacity); us-east-1 currently has no g6e capacity.
python vlasim.py deploy --vla molmoact2 --email you@example.com --region us-west-2

# RLDX-1 (RLWRLD) — LIBERO (SOTA open VLA, MSAT/Qwen3-VL-8B, ZeroMQ server + sim client, ~80-120 min)
# NON-COMMERCIAL weights (sim benchmarking) — AWS-internal enablement only.
python vlasim.py deploy --vla rldx --email you@example.com

# RLDX-1 (RLWRLD) — SimplerEnv Google-VM (real-robot OXE_FRACTAL embodiment, ~35-45 min)
# NON-COMMERCIAL weights (sim benchmarking) — AWS-internal enablement only.
python vlasim.py deploy --vla rldx-simpler --email you@example.com

# RLDX-1 (RLWRLD) — RoboCasa GR-1 Tabletop (bimanual humanoid + waist, MuJoCo, ~25-35 min)
# NON-COMMERCIAL weights (sim benchmarking) — AWS-internal enablement only.
python vlasim.py deploy --vla rldx-gr1 --email you@example.com

# RLDX-1 (RLWRLD) — RoboCasa Kitchen 24-task (PandaOmron mobile manipulator, MuJoCo, ~25-35 min)
# NON-COMMERCIAL weights (sim benchmarking) — AWS-internal enablement only.
python vlasim.py deploy --vla rldx-kitchen --email you@example.com

# OpenArm Lift-Cube — scripted ACT demo collection in Isaac Lab (HDF5 only; needs a 4-GPU instance)
python vlasim.py deploy --vla openarm-lift-act --email you@example.com

On first deploy you will receive an SNS subscription confirmation email — click the link to enable notifications.

Direct invocation. vlasim deploy is a thin wrapper — every command above also runs as python deploy.py --vla <target> ... with the same flags, skipping the doctor preflight. Use that if you've already run doctor and want to bypass it (equivalently, vlasim deploy --skip-doctor).

Monitor (optional)

# GR00T N1.7 logs
aws logs tail /gr00t/userdata --follow --region us-east-1

# GR00T N1.6 + GR1 logs
aws logs tail /gr00t-gr1/userdata --follow --region us-east-1

# GR00T N1.6 + Unitree G1 logs
aws logs tail /gr00t-g1/userdata --follow --region us-east-1

# π0.5 logs
aws logs tail /pi/userdata --follow --region us-east-1

# OpenVLA-OFT logs
aws logs tail /openvla-oft/userdata --follow --region us-east-1

# LAP-3B logs
aws logs tail /lap/userdata --follow --region us-east-1

# RLDX-1 logs
aws logs tail /rldx/userdata --follow --region us-east-1

# RLDX-1 SimplerEnv logs
aws logs tail /rldx-simpler/userdata --follow --region us-east-1

# RLDX-1 RoboCasa GR-1 Tabletop logs
aws logs tail /rldx-gr1/userdata --follow --region us-east-1

# RLDX-1 RoboCasa Kitchen logs
aws logs tail /rldx-kitchen/userdata --follow --region us-east-1

Results

When simulation finishes you receive an SNS email with download instructions:

# GR00T results
aws s3 sync s3://vla-sim-results-gr00t-demo-us-east-1-<ACCOUNT>/RUN_ID/ ./results/ --region us-east-1

# π0.5 results
aws s3 sync s3://vla-sim-results-pi-demo-us-east-1-<ACCOUNT>/RUN_ID/ ./results/ --region us-east-1

Each task-N/ folder contains:

videos/rollout_<task_name>_success.mp4 — successful episodes
videos/rollout_<task_name>_failure.mp4 — failed episodes
videos/<name>.txt — per-video metadata
summary.txt — success_rate, suite, num_episodes_per_task
compose.log / userdata.log — execution logs

Cleanup

python vlasim.py destroy --vla gr00t
python vlasim.py destroy --vla gr00t-gr1
python vlasim.py destroy --vla gr00t-g1
python vlasim.py destroy --vla pi
python vlasim.py destroy --vla openvla-oft                         # default suite (10)
python vlasim.py destroy --vla openvla-oft --libero-suite spatial  # non-default suite
python vlasim.py destroy --vla lap
python vlasim.py destroy --vla rldx
python vlasim.py destroy --vla rldx-simpler
python vlasim.py destroy --vla rldx-gr1
python vlasim.py destroy --vla rldx-kitchen

vlasim destroy forwards to destroy.py unchanged — python destroy.py --vla <target> works identically if you prefer to call it directly.

The S3 results bucket is retained after stack deletion to preserve simulation outputs.

Expected Results

`--vla`	Task	Expected Success Rate	Source
`gr00t`	KITCHEN_SCENE3 (stove + moka pot)	~90–100%	validated 2026-04-27
`gr00t`	KITCHEN_SCENE4 (black bowl in drawer)	~90–100%	validated 2026-04-27
`gr00t-gr1`	PosttrainPnPNovelFromPlateToBowlSplitA (GR1)	~80%	validated 2026-04-12
`gr00t-gr1`	PnPCanToDrawerClose (GR1)	0% (pre-trained N1.6 not supported)	validated 2026-04-12
`gr00t-g1`	LMPnPAppleToPlateDC (G1 loco-manip, community re-finetune)	30% (3/10)	validated 2026-06-15 — `nvidia/...` checkpoint scores 0%; `cloudwalk-research` re-finetune used
`pi`	libero_object	~80–94%	validated 2026-04-27
`pi`	libero_spatial	~85–95%	validated 2026-04-27
`openvla-oft --libero-suite spatial`	libero_spatial	97.6% (paper Table I)	validated 2026-06-01 — 0.92 (46/50, 5 trials/task × 10 tasks, g6.xlarge); within paper ±5%p band at n=50
`openvla-oft --libero-suite object`	libero_object	98.4% (paper Table I)	pending smoke test
`openvla-oft --libero-suite goal`	libero_goal	97.9% (paper Table I)	pending smoke test
`openvla-oft --libero-suite 10`	libero_10 (long-horizon)	94.5% (paper Table I)	validated 2026-05-04
`lap`	libero_spatial	~85-95% (paper Table III, LIBERO fine-tuned)	validated 2026-05-17 — 0.98 @ 5 trials/task
`molmoact2`	libero_goal (2 tasks × 5 ep)	97.2% LIBERO avg (paper)	validated 2026-06-28 — 1.000 (10/10 ep, fp32 in-process, g6e.2xlarge L40S 48GB, full `deploy.py` 1-shot smoke, us-west-2)
`rldx`	libero_spatial (single task)	97.8% LIBERO avg (paper, SOTA)	validated 2026-06-20 — 1.0 (5/5 ep, eager, g6.xlarge L4 sm_89, full `deploy.py` smoke)
`rldx-simpler`	simpler_env_google/google_robot_pick_coke_can (Google-VM Fractal)	81.5% Google-VM (paper README)	validated 2026-06-24 — 1.0 (5/5 ep, eager, g6.2xlarge L4, full `deploy.py` 1-shot smoke)
`rldx-simpler`	simpler_env_google/google_robot_move_near (Google-VM Fractal)	81.5% Google-VM (paper README)	validated 2026-06-24 — 1.0 (5/5 ep, eager, g6.2xlarge L4, full `deploy.py` 1-shot smoke)
`rldx-gr1`	gr1_unified/PnPCupToDrawerClose (GR-1 bimanual + waist)	58.7% 24-task avg (paper README)	validated 2026-06-25 — 0.80 (4/5 ep, eager, g5.2xlarge A10G 24GB, full `deploy.py` 1-shot smoke)
`rldx-gr1`	gr1_unified/PnPMilkToMicrowaveClose (GR-1 bimanual + waist)	58.7% 24-task avg (paper README)	validated 2026-06-25 — 0.20 (1/5 ep, eager, g5.2xlarge A10G 24GB, full `deploy.py` 1-shot smoke)
`rldx-kitchen`	robocasa_panda_omron/OpenSingleDoor (PandaOmron mobile manip)	70.6% 24-task avg (paper README)	0.80 (4/5), validated 2026-06-26 (g6.2xlarge L4)
`rldx-kitchen`	robocasa_panda_omron/OpenDrawer (PandaOmron mobile manip)	70.6% 24-task avg (paper README)	0.80 (4/5), validated 2026-06-26 (g6.2xlarge L4)

Validated results (us-east-1, g6.12xlarge / g5.xlarge / g6.xlarge):

GR00T N1.7: KITCHEN_SCENE3 = 1.0 (5/5), KITCHEN_SCENE4 = 1.0 (3/3)
GR00T N1.6 + GR1: PosttrainPnP = 0.8 (4/5), PnPCanToDrawer = 0.0 (pre-trained model limitation)
GR00T N1.6 + Unitree G1 (whole-body loco-manip): LMPnPAppleToPlateDC = 0.3 (3/10, n_envs=1, g6.12xlarge, us-west-2) — cloudwalk-research community re-finetune; the nvidia/GR00T-N1.6-G1-PnPAppleToPlate checkpoint scores 0/10 under the same command (matches open Isaac-GR00T issues, no upstream fix)
π0.5: libero_object = 0.94 (47/50)
OpenVLA-OFT: libero_10 = 1.0 (10/10 at 1 trial/task, g6.xlarge)
OpenVLA-OFT: libero_spatial = 0.92 (46/50, 5 trials/task × 10 tasks, g6.xlarge) — paper Table I = 97.6%; the gap (5.6%p) is within sampling noise at n=50 (SE ≈ 3.8%p). 8/10 tasks at 5/5; misses concentrated on on_the_ramekin (2/5) and next_to_the_plate (4/5)
LAP-3B: libero_spatial = 0.98 (49/50, 5 trials/task × 10 tasks, g6.xlarge) — exceeds paper Table III range (85–95%); requires upstream scripts/libero/main.py vertical-flip patch (see templates/lap-userdata.sh.j2)
MolmoAct2 (Ai2): libero_goal = 1.000 (10/10, 2 tasks × 5 ep, g6e.2xlarge L40S 48GB, us-west-2) — paper LIBERO average = 97.2%. Apache-2.0 5B Action Reasoning Model run in-process on the sim GPU (no policy server) from the allenai/lerobot fork (molmoact2-policy@a4f15bf3, not upstream); official fp32 LIBERO recipe (model_dtype=float32, use_amp=false, norm_tag=libero, continuous). The ~20 GB fp32 checkpoint forces a ≥40 GB single GPU and ≥64 GiB host RAM (g6e.xlarge's 32 GiB host RAM thrashes the load → excluded); checkpoint allenai/MolmoAct2-LIBERO@0d24a92b, Python pinned 3.12. Full deploy.py 1-shot smoke; pipe-proof cap was 2 tasks × 5 ep (drop task_ids + n_episodes: 50 for the full 10-task suite)
RLDX-1 (RLWRLD): libero_10 long-horizon, two tasks (eager, g6.xlarge L4 sm_89) — STUDY_SCENE1 book→caddy = 1.0 (5/5, tight 20–23 steps), KITCHEN_SCENE6 mug→microwave+close = 0.8 (4/5; successes span 32–47 steps and the one failure is a 90-step-cap timeout — marginal even when it succeeds, see the showcase caveat). Full deploy.py deploy; ZeroMQ policy server + sim client, two-venv (main + libero_uv). Default n_envs=1 (5 recommended on g6.2xlarge+); checkpoint RLWRLD/RLDX-1-FT-LIBERO@989037c6, repo RLWRLD/RLDX-1@ecbfaf80
RLDX-1 (RLWRLD) GR-1 Tabletop: two articulated tasks (eager, g5.2xlarge A10G 24GB) — PnPCupToDrawerClose = 0.80 (4/5, clean 13–19 steps), PnPMilkToMicrowaveClose = 0.20 (1/5; the single success at 24 steps, the four failures all hit the 45-step cap as timeouts). Combined 5/10 ≈ paper 24-task avg 58.7%. The only target on a bimanual humanoid + waist (GR1ArmsAndWaistFourierHands); MuJoCo/robosuite headless EGL (no Vulkan). Source-verified needed no SimplerEnv-style code patch — confirmed live: the [5/8] gr1-venv verify (dispatch + obs/action contract + import-time env registration) passed without FIX 4/6/7, and setup_gr1.sh exited 0 (committed-as-files submodule → git submodule update is a no-op, no FIX 5). Server cold-load fast (~210s). Checkpoint RLWRLD/RLDX-1-FT-GR1@d0277c5c, repo RLWRLD/RLDX-1@ecbfaf80

Configuration Reference

`simulator-config.yaml` (shared)

Key	Default	Description
`deployment.region`	`us-east-1`	AWS region
`deployment.notify_email`	(required)	SNS notification email
`deployment.s3_results_prefix`	`vla-sim-results`	S3 bucket name prefix
`deployment.auto_terminate`	`true`	Auto-terminate EC2 after completion

`models/gr00t.yaml`

Key	Description
`model.hf_repo`	HuggingFace repo for checkpoint download
`model.hf_subfolder`	Subfolder within repo (e.g. `libero_10`)
`model.hf_model_revision`	Pinned commit SHA for reproducibility
`model.isaac_groot_commit`	Pinned Isaac-GR00T commit SHA
`instance.preferred`	GPU instance type fallback list
`instance.ebs_gb`	Root EBS volume size (GB)
`tasks`	List of LIBERO tasks to evaluate

`models/pi.yaml`

Key	Description
`model.openpi_commit`	Pinned openpi commit SHA
`instance.preferred`	GPU instance type fallback list
`tasks[].suite`	LIBERO suite name (`libero_spatial`, `libero_object`, etc.)
`tasks[].num_episodes`	Episodes per task

Project Structure

vla-simulator/
├── vlasim.py                 # CLI front end: doctor / init / deploy / destroy
├── simulator-config.yaml     # Shared deployment settings
├── docs/
│   └── showcase/             # Sample rollout MP4 + GIF preview per VLA policy
├── models/
│   ├── gr00t.yaml            # GR00T N1.7 config
│   ├── gr00t-gr1.yaml        # GR00T N1.6 + GR1 humanoid config
│   ├── gr00t-g1.yaml         # GR00T N1.6 + Unitree G1 whole-body loco-manip config
│   ├── pi.yaml               # π0.5 config
│   ├── openvla-oft.yaml      # OpenVLA-OFT config (per-suite checkpoints)
│   └── lap.yaml              # LAP-3B config
├── deploy.py                 # 1-click deploy entrypoint
├── destroy.py                # Stack teardown
├── generate.py               # Generates assets/userdata/{vla}.sh from templates
├── requirements.txt
├── cdk/
│   ├── bin/app.ts
│   └── lib/
│       ├── vla-simulator-stack.ts   # Unified CDK stack
│       └── az-selector.ts           # GPU capacity Lambda (AZ/type fallback)
├── assets/
│   ├── userdata/             # Generated scripts (gitignored)
│   └── bridge/
│       ├── gr00t/            # ZMQ-gRPC bridge for GR00T
│       ├── pi/               # gRPC bridge for π0.5
│       └── lap/              # WebSocket↔gRPC bridge for LAP-3B (port 50055)
└── templates/
    ├── gr00t-userdata.sh.j2        # Jinja2 template for GR00T UserData
    ├── gr00t-gr1-userdata.sh.j2    # GR00T N1.6 + GR1 humanoid
    ├── gr00t-g1-userdata.sh.j2     # GR00T N1.6 + Unitree G1 (GR00T-WholeBodyControl loco-manip)
    ├── pi-userdata.sh.j2           # π0.5 (Docker Compose)
    ├── openvla-oft-userdata.sh.j2  # OpenVLA-OFT (single conda env)
    ├── lap-userdata.sh.j2          # LAP-3B (uv 2-venv: JAX policy + LIBERO sim)
    ├── openarm-isaac-userdata.sh.j2     # OpenArm bimanual π0.5 eval (Isaac Lab container)
    └── openarm-lift-act-userdata.sh.j2  # OpenArm Lift-Cube scripted ACT collection (Isaac Lab)

Bridge Mode

Bridge mode connects the simulation environment on EC2 to an external VLA model endpoint (e.g. VLA Hub on ECS), without downloading the model locally.

GR00T bridge

# Set endpoint in models/gr00t.yaml (or SSM):
#   bridge.remote_grpc_endpoint: ssm:/vla-hub/gr00t/n1-7/grpc-endpoint
#   bridge.vpc_id: ssm:/vla-hub/vpc-id

python vlasim.py deploy --vla gr00t --bridge

π0.5 bridge

# Set in models/pi.yaml:
#   bridge.vpc_id: vpc-xxxxxxxxxxxxxxxxx
#   bridge.nlb_endpoint: internal-xxx.elb.us-east-1.amazonaws.com:50052

python vlasim.py deploy --vla pi --bridge

LAP-3B bridge

LAP shares π0.5's openpi WebSocket bridge pattern. The instance runs only the LIBERO sim + a WebSocket↔gRPC bridge (assets/bridge/lap/); the LAP-3B JAX model runs on the remote vla-hub LAP ECS task (port 50055). No local checkpoint download.

# Set in models/lap.yaml (or use ssm: for nlb_endpoint):
#   bridge.vpc_id: vpc-xxxxxxxxxxxxxxxxx                       # the vla-hub VPC
#   bridge.nlb_endpoint: internal-xxx.elb.us-west-2.amazonaws.com:50055
#                        or ssm:/vla-hub/lap/3b/grpc-endpoint

python vlasim.py deploy --vla lap --bridge

Bridge differs from π0.5 only in the wire contract: LAP uses a nested observation (observation.base_0_rgb / left_wrist_0_rgb / state), a 10-dim state (eef_pos 3 + eef_rot6d 6 + gripper 1), a frame_description field, and a (10,7) action chunk — see assets/bridge/lap/lap.proto.

Troubleshooting

"Embodiment tag 'LIBERO_PANDA' is not supported"

The base GR00T N1.7-3B model does not support LIBERO simulation. Use the fine-tuned checkpoint:

# models/gr00t.yaml
model:
  hf_repo: nvidia/GR00T-N1.7-LIBERO
  hf_subfolder: libero_10

"No visible GPU devices" (π0.5 task-0)

JAX inside the Docker container fails if GPU passthrough isn't ready before the task loop starts. The [2.5/5] step in the UserData script polls docker run --gpus all nvidia-smi until GPU is confirmed before proceeding.

Stack rollback / InsufficientInstanceCapacity

The AzSelectorConstruct Lambda automatically retries across AZs and falls back through the instance type preference list (g6.12xlarge → g5.12xlarge → g6.xlarge → g5.xlarge for GR00T). No manual intervention needed — CDK rollback means no capacity was found; re-deploying will retry.

CDK synth cdk.out conflict (parallel deploys)

Use model-specific output directories:

npx cdk deploy GR00T-Demo --output cdk.out-gr00t -c vla=gr00t -c region=us-east-1 ...
npx cdk deploy Pi-Demo    --output cdk.out-pi    -c vla=pi    -c region=us-east-1 ...

deploy.py handles this automatically.

apt lock on DLAMI boot

unattended-upgrades holds the apt lock for ~2 minutes on fresh boot. The UserData waits up to 120 seconds then forcibly kills the process. No action needed.

Cost Estimate (us-east-1, on-demand)

Instance	Spot/OD	Hourly	GR00T (~2h)	π0.5 (~4h)	LAP (~2.5h)
g6.12xlarge	On-Demand	~$4.09	~$8.18	—	—
g5.xlarge	On-Demand	~$1.01	—	~$4.04	—
g6.xlarge	On-Demand	~$0.80	—	—	~$2.00

Use Spot instances via deploy.py --spot (not yet implemented) for 60–70% savings.

Related Projects

Isaac-GR00T — NVIDIA GR00T foundation model
openpi — Physical Intelligence π0.5
openvla-oft — OpenVLA-OFT (Optimized Fine-Tuning)
lap — LAP: Language-Action Pre-Training (zero-shot cross-embodiment, JAX)
RLDX-1 — RLWRLD dexterity-first VLA (MSAT / Qwen3-VL-8B, Apache-2.0 code)
LIBERO — Long-horizon robot benchmark
SimplerEnv + ManiSkill2_real2sim — real-to-sim eval for Google Fractal / WidowX Bridge (MIT)
robocasa-gr1-tabletop-tasks — RoboCasa GR-1 humanoid tabletop benchmark (24 tasks, MIT)

Acknowledgments & Licensing

This sample's own code (CDK, generators, templates) is MIT-0 (see LICENSE). It orchestrates several third-party models and simulators, each under its own license — pulled at runtime (not vendored in this repo):

Component	Used by	License	Notes
NVIDIA Isaac-GR00T (N1.7 / N1.6)	`gr00t*`	NVIDIA license	weights from Hugging Face
Physical Intelligence openpi (π0.5)	`pi`, `openarm-isaac`	Apache-2.0
OpenVLA-OFT	`openvla-oft`	MIT
LAP-3B	`lap`	MIT (code)
RLDX-1 (RLWRLD)	`rldx`, `rldx-simpler`, `rldx-gr1`, `rldx-kitchen`	code Apache-2.0; weights RLWRLD Model License v1.0 (NON-COMMERCIAL)	"simulation benchmarking" is an explicit intended use. AWS-internal enablement / benchmarking only — not for customer-facing commercial positioning.
LIBERO	LIBERO targets	MIT
SimplerEnv + ManiSkill2_real2sim	`rldx-simpler`	MIT	cloned at runtime by `setup_SimplerEnv.sh`
RoboCasa GR-1 Tabletop tasks	`rldx-gr1`	MIT	cloned at runtime by `setup_gr1.sh` (`robocasa/robocasa-gr1-tabletop-tasks`)
RoboCasa (Kitchen) + robosuite	`rldx-kitchen`	MIT	cloned at runtime by `setup_robocasa.sh` (`squarefk/robocasa`, fork of ARISE-Initiative/robocasa); ~5 GB scene assets download during setup

The rldx/rldx-simpler/rldx-gr1/rldx-kitchen targets are built on RLDX-1 by RLWRLD — gratefully acknowledged. RLDX-1's eval harness vendors LIBERO, SimplerEnv, RoboCasa, and GR-1 task suites as submodules (all MIT); this sample clones the RLDX-1 repo at a pinned commit at runtime and applies small in-place runtime fixes to the clone only — the upstream repo is never modified and no model weights are committed here. Per target: rldx needs only the shared dependency-pin fixes; rldx-simpler adds a SimplerEnv dispatch + obs/action adapter patch; rldx-gr1 needs no code patch (dispatch + obs/action routing already correct at the pin); rldx-kitchen needs no dispatch/obs/action patch but adds three RoboCasa-Kitchen-specific runtime fixes (making the cloned asset downloader non-interactive for headless setup; a numpy/numba pin reconciliation so the import-reachable albumentations resolves against the kitchen fork; and a cosmetic video-recorder narrowing so the saved clip shows each camera once rather than duplicating it at two resolutions).

Name		Name	Last commit message	Last commit date
Latest commit History 14 Commits
assets		assets
cdk		cdk
docs/showcase		docs/showcase
models		models
templates		templates
.gitignore		.gitignore
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
NOTICE		NOTICE
README.md		README.md
deploy.py		deploy.py
destroy.py		destroy.py
generate.py		generate.py
requirements.txt		requirements.txt
simulator-config.yaml		simulator-config.yaml
vlasim.py		vlasim.py

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

VLA Simulator — 1-Click VLA Simulation on AWS

Overview

Supported VLA Combinations

Showcase — VLA Rollouts

pi — π0.5 on LIBERO (Franka Panda 7-DOF)

openvla-oft — OpenVLA-OFT on LIBERO-10 long-horizon

lap — LAP-3B on LIBERO-Spatial

molmoact2 — MolmoAct2 (Ai2) on LIBERO-Goal (Franka Panda 7-DOF)

rldx — RLDX-1 (RLWRLD) on LIBERO-10 long-horizon (Franka Panda 7-DOF)

rldx-simpler — RLDX-1 (RLWRLD) on SimplerEnv Google Fractal (real-robot embodiment)

rldx-gr1 — RLDX-1 (RLWRLD) on RoboCasa GR-1 Tabletop (bimanual humanoid + waist)

rldx-kitchen — RLDX-1 (RLWRLD) on RoboCasa Kitchen (PandaOmron mobile manipulator)

gr00t-gr1 — GR00T N1.6 on RoboCasa GR1 humanoid (22-DOF)

gr00t-g1 — GR00T N1.6 on Unitree G1 (whole-body loco-manipulation)

Optional: run a fine-tuned GR00T N1.7 Unitree-G1 adapter on the same sim

gr00t — GR00T N1.7 on LIBERO-10 (Franka Panda)

openarm-lift-act — OpenArm unimanual Lift-Cube (Isaac Lab)

Architecture

Quick Start

Deploy Targets

Monitor (optional)

Results

Cleanup

Expected Results

Configuration Reference

simulator-config.yaml (shared)

models/gr00t.yaml

models/pi.yaml

Project Structure

Bridge Mode

GR00T bridge

π0.5 bridge

LAP-3B bridge

Troubleshooting

"Embodiment tag 'LIBERO_PANDA' is not supported"

"No visible GPU devices" (π0.5 task-0)

Stack rollback / InsufficientInstanceCapacity

CDK synth cdk.out conflict (parallel deploys)

apt lock on DLAMI boot

Cost Estimate (us-east-1, on-demand)

Related Projects

Acknowledgments & Licensing

About

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages