ManiSkill scene evidence
Explore one Panda cube-pickup scene with two recordings: an official motion-planning demonstration and a locally executed released PPO policy.
They use different pinned ManiSkill versions and controllers. Reset states and policy recordings do not add distinct worlds.
Source and attribution
ManiSkill contributors: pinned upstream source, ManiSkill 3.0.1, SAPIEN 3.0.3 and PhysX CUDA. The upstream licence notice assigns Apache 2.0 to code and CC BY-NC 4.0 to assets.
Meshes and textures load from their original, commit-pinned URLs. This page does not grant commercial asset rights.
The derived display description retains native visual transforms, original mesh scales, and nine independently recorded Panda joints.
Finger mimic is removed from the display description to preserve native finger asymmetry. The procedural ground exports native triangles and texture coordinates.
Browser rendering adapts lighting and camera coordinates; it does not execute SAPIEN physics.
Motion-planning demonstration · 3.0.1
PickCube-v1, seed 0, official motion-planning solution: 82 actions and 83 states including reset. The original episode ends at step 50; the first recorded goal success is step 77.
Later completion is shown as demonstration data and receives no in-budget success credit. No learned checkpoint, suite score, contact force or reward stream is claimed.
The controller accepts seven absolute arm joint targets in radians and one normalized gripper target. Replay uses recorded joint and body states, not action execution.
ManiSkill 3 remains distinct from historical ManiSkill2 results.
PPO reference evaluation · 3.0.0b15 · 2026-09-13
The released PickCube PPO checkpoint was executed through the original reference evaluator at ManiSkill commit baab60ede2e89167c1b7aaed41a9aa8e690a9d1e.
Its accompanying source trajectory metadata identifies this older 3.0.0b15 version.
The separate runtime used Python 3.10.20, SAPIEN 3.0.0b1, PyTorch 2.4.1+cu124 and an NVIDIA RTX 4090. The checkpoint's original training package versions are unpublished.
Ten reference seeds, 1 through 10, were selected before execution. Each completed its original 50-action episode twice in separate native processes: 10/10 success at the end in each set.
These are two executions of the same ten test conditions, not twenty independent conditions. The browser recording uses the first selected seed, which first reaches success at action 31.
This result covers one task and one released checkpoint; it is not a full ManiSkill suite or a reproduction of a paper score.
Every policy action matches independent NumPy checkpoint inference within 4.84e-7; every 42-dimensional policy observation matches native state within 1.31e-8.
The task succeeds when the cube is within 0.025 m of the goal and the maximum absolute arm joint velocity is at most 0.2, excluding the two finger velocities.
All source-owned HDF5 state/action/observation/reward/outcome datasets match the observer, and each independent pair has identical dataset values.
The original evaluator ignores success termination signals until action 50; raw native termination flags remain in the recording.
The policy supplies eight raw outputs to pd_joint_delta_pos; outputs can exceed the controller's declared [-1,1] Box. Original upstream processing is retained.
Replay applies the observed body world poses directly and preserves joint telemetry. Restoring joint positions alone differed by up to 2.57 micrometres at a finger, so the original observed link poses supply the replay.
Desktop component checks compared 3,672 points against original moving/static recorded body poses with maximum error below 7.29e-8 m. This component test is not a live-network performance or synchronized-video claim.
The checkpoint dataset card declares Apache 2.0; the pinned source notice separately assigns CC BY-NC 4.0 to assets. Weights are not website assets.
The initial observer attempt stopped before any action because its horizon lookup did not account for the source wrapper; its failed receipt is retained with the native fixtures.
The corrected observer uses the original source's horizon lookup without changing the task, seeds, controller or physics.