Newton native FR3 replay
Watch a Franka Research 3 arm lift a pen and place it in a cup. The original Newton example computes joint targets from scripted waypoints and inverse kinematics. Its source filename says “Panda hydro,” but the loaded robot is FR3.
Open recording · Native verification · Reproduction notes (repository access required)
What the recording establishes
| Item | Recorded evidence |
|---|---|
| Example | Newton 1.6.0, pen scene, one world. Original FR3, pad and cup inputs; original camera and scripted IK controller. Deterministic contact ordering and deterministic solver enabled. |
| Lift | Maximum pen lift 0.210853 m; source threshold >0.15 m. |
| Final placement checks | Pen/cup planar offset 0.016392 m (<0.02 m); cup up-axis z approximately 1 (>0.95); pen origin 0.035023 m above cup origin (>−0.05 m). These are the upstream geometric predicates. |
| Independent replay | A fresh native model applies recorded controls with no per-step IK controller calls. All 721 samples and 181,692 checked joint/body state scalars match exactly; the upstream final checks pass again. This establishes repeatability on the tested host, not across hardware. |
| Video | 721 native RGB frames, 640 × 480, 60 fps. H.264 delivery: 1,393,957 bytes; 12.016667 s. Every decoded frame compared with its native input; minimum PSNR 39.4473 dB. The original view shows the manipulation area and crops part of the upper arm. |
| Browser | One simulation recording with camera/state playback, seeking, metadata and recipe downloads. Zero browser world assemblies, learned checkpoints or reproduced evaluation suites. |
Time, state and controls
Sample 0 is the post-constructor state, with no issued step command. Each subsequent sample follows one controller command and ten native physics substeps of 1/600 s. Encoded frame i has presentation time i/60 s. Measured native time is retained; maximum discrepancy is below 2e-13 s. The final state at approximately 12 s occupies the last 1/60-second frame.
The recording ends at the source example’s 720-step frame budget. The metadata marks recording_truncated=true at the final sample. No Gym reward, termination or truncation signal exists; the required terminated/truncated fields stay false and do not assert an upstream environment signal.
Downloads retain 23 joint-position coordinates, 21 generalized velocities and 16 body poses. The robot has seven arm hinges and two finger slides; the pen and cup each have a free joint. Nine absolute robot position targets form the action vector. Full native target-position, target-velocity and force arrays remain separate. No learned policy or Gym action bounds are inferred.
Native right-handed z-up positions are unchanged. Body and free-joint quaternion fields change order from xyzw to wxyz; original arrays remain in the raw native samples. Generalized velocities keep native order: seven arm rad/s, two finger m/s, then pen and cup linear m/s and angular rad/s in their joint parent frames. Free-joint linear velocities refer to child centres of mass. Raw body velocities retain world-frame linear COM velocity followed by angular velocity, as documented in the pinned State implementation.
The coordinate audit compares all 11,536 body transforms against Warp 1.17.0, using 46,144 transformed basis/origin points. Maximum absolute error: 1.363e-7. Both free-joint poses match the corresponding body poses at all 721 samples. Original float quaternions retain their small norm errors. No camera matrix, 3D overlay or converted browser geometry is fabricated.
Inputs and reproduction
The Newton source revision c2ca70be is extracted from exact Git objects: 1,150 regular files / 34,307,252 bytes. The selected asset folders at newton-assets f8fb7abc contain 72 regular files / 81,108,018 bytes. The input receipt records Git identities, modes, sizes and SHA-256 hashes.
The recursive asset audit resolves the selected FR3 URDF meshes, COLLADA image references and USD layer/asset references: 24 files / 27,841,296 bytes, with zero unresolved dependencies and unchanged input bytes. Original robot and object geometry stays in native storage.
The isolated runtime uses Python 3.12.3, Newton 1.6.0, Warp 1.17.0, MuJoCo Warp 3.12.0 and OpenUSD 26.3 on an RTX 4090 with NVIDIA driver 580.173.02. The evidence binds the 113 installed package versions and hashed dependency lock. This standalone environment is separate from the existing Isaac drawer’s Newton 1.5.0 runtime.
The upstream sources are public. The exact capture, replay and publishing scripts are in this site’s private repository; recipe links require repository access. The downloadable recipe names the site build commit, selected inputs and commands for a fresh native trial. A new run must be qualified before replacing this evidence.
Attribution and limits
Newton contributors provide the engine under Apache-2.0. The selected Franka robot, pad and cup folders carry Apache-2.0 notices. The robot README traces its original description to Franka Robotics. Asset authors retain their rights; this site generated the rendered video and measured-state recording.
This is one scripted manipulation example. The source checks do not establish contact-based insertion accuracy, real-robot performance, a learned-policy result or a benchmark suite score. No original URDF, mesh, USD or engine distribution is rehosted here. The frontier sequence tracks the remaining Genesis, aerial and marine candidates.