curobo

cuRobo V2 motion planning

The image candidate remains 0.8.0-cuda13-b300-unbuilt and publication-quarantined until built-image checks and real GPU validation pass. Build from committed inputs; build.sh checks scoped source cleanliness and archives the exact commit for Docker. The tag family does not establish B300 validation.

For image changes, inspect actual dependency wheels and base layers. The cuDNN development base carries headers; deleting inherited files in a later layer does not remove those bytes. This image uses a non-cuDNN CUDA base and filters the locked cuDNN wheel inside its installation RUN, retaining only shared runtime libraries and notices. The bundled older and current cuDNN supplements differ on headers; preserve both evidence and use the common runtime boundary. Keep the filter's version/inventory checks, then inspect every built layer before publication. Passing its source tests does not establish clean image bytes.

Run the GPU workflow on Nebius through the ordinary workflow submit path:

npa workbench workflow validate-spec workflows/testing/curobo-benchmark.yaml
npa workbench workflow submit workflows/testing/curobo-benchmark.yaml --var bucket=<your-bucket>

This is the complete benchmark: both pinned datasets, in kinematic and 3 kg payload dynamics modes. curobo_mode=kinematic|dynamics selects one complete configuration. Do not silently replace either dataset with the upstream demo or add problem/time/job limits. The golden evaluation separately qualifies one real pose; it does not prove full-benchmark completion.

Exact implementation and limits

• Source is cuRobo V2 at 8e734f3ced1df898990bcd92de40abce475907db, using MotionPlanner and MotionPlannerCfg. V1 MotionGen examples are incompatible.
• Raw benchmark datasets are robometrics 81e3d1d605de84100d8ab880b43096aba221a48b. V2 source and Franka assets are Apache-2.0; dataset/source MIT and MotionBenchMaker BSD notices are retained. No weights, gated access or model-acceptance switch is required.
• The benchmark calls upstream's configuration loader, including its relaxed joint limits (0.2 radians), obstacle-to-OBB conversion and optimizer settings. Negative collision_buffer_ik inputs remain recorded as invalid.
• Every input contributes to the full denominator. Eligible success is reported separately. Failed solves remain failed; inverse-dynamics errors fail the job instead of becoming zero energy. FK path lengths are computed from actual tool positions, not upstream's placeholder end-effector metrics.
• A matching total count is insufficient evidence. Validate exact problem identities and invalid indices against benchmark_inventory.py, independently derived from the pinned YAML with file hashes. The runner checks those bytes, and report validation requires the known metrics for every status plus sample timeline consistency. Do not accept a self-consistent but invented journal.
• Energy is a Pinocchio inverse-dynamics proxy on the optimized joint trajectory. Planner success is upstream feasibility, not independent collision certification or authorization to move physical hardware.