Isaac Lab native drawer replay
Watch a Franka Panda open a cabinet’s top drawer using an official RSL-RL policy checkpoint. This is one native task recording, with the camera and measured states on the same timeline.
Open the recording · Inputs and verification · Reproduction recipe
What the recording establishes
| Item | Recorded evidence |
|---|---|
| Task | Isaac-Open-Drawer-Franka; one environment; seed 0; source development revision ddb54d34. |
| Native result | The source success flag first becomes true at step 22, or 0.3667 seconds. Maximum drawer opening: 0.4101 m; success threshold: greater than 0.30 m. |
| Replay proof | A second native execution replays the recorded actions without loading the policy checkpoint. All 481 samples and 79,846 joint/velocity/body-pose values match exactly, along with reward and outcome flags. |
| Video | 481 native RGB frames, 640 × 480, 60 fps. Encoded clip: 584,957 bytes; all frames compared with native RGB, minimum PSNR 45.93 dB. |
| Browser capabilities | Camera/state playback, seeking, metadata and reproduction recipe. One native successful task; zero browser USD world assemblies. Native CUDA execution is required to reproduce the run. |
Time and state
Sample 0 is the reset state, without an action or reward. Sample i greater than zero contains the state after action i and reward computation, before automatic reset. Control time and encoded presentation time both equal i/60 seconds. The last state is at 8 seconds; its final video frame ends at 8.0167 seconds.
The source task ends on its 8-second timeout, even after success. Final flags are terminated=false, truncated=true, success=true. The recording retains all 13 joint positions and velocities, 20 body poses, 8 action inputs, reward, camera pose/intrinsics and native task flags. The plot shows Franka joints 1 and 2 in radians. Body quaternions are explicitly reordered from native xyzw to the common wxyz export contract; native camera data retains its stated xyzw convention.
Native runtime and limits
The source pin uses Newton 1.5.0, mujoco-warp 3.11.0 and Warp 1.16.0. The observed host uses an RTX 4090 with driver 580.173.02 and Python 3.12.13. This is a development revision and an existing runtime, not a fresh installation or a compatibility claim for all Isaac Lab or Arena tasks.
The observer camera is outside the policy’s 31 state inputs but is part of the exact native configuration. It changes the included visual shapes and the task’s per-shape friction randomization. Camera-free replay diverges; restoring the common friction values alone does not remove later divergence. The matching-camera configuration is the verified one.
The USD dependency audit resolves 23 vendor layer/texture files plus the generated root layer, while OmniPBR.mdl remains unresolved. The selected Newton renderer uses authored USD colors through the pinned source’s material adapter. The images do not reproduce Kit/MDL shading. Original USD, geometry and physics stay native; no converted browser scene is supplied.
The official task/backend checkpoint is identified by its URL, size and SHA-256. Its training-source revision is unknown. This single successful trial is not a reproduced suite score, an Arena composition or a Galbot hardware demonstration.
Source and output terms
Isaac Lab code is BSD-3-Clause. Its pinned Franka asset notice is Apache-2.0. Other NVIDIA assets have additional materials terms. NVIDIA’s licence FAQ addresses publication of simulation outputs. This release publishes measured state and rendered video without redistributing the original assets, runtime or policy weights.