DVIDIA RESEARCH · NATIVE CONTACT TASK
Grasp. Carry. Release. Measure.
Authored, privileged-state rigid-block task in native reference physics. No learned policy, shoe physics, human-video action recovery or physical robot qualification.
A free rigid block must be lifted, placed near a separate target, released and left settled with the tool retreated. The scorer uses native object state independently of the controller’s phase. The block is never welded to the gripper.
Full summary · Frozen protocol · Artifact hashes
Every planned attempt
18 attempts · confirmation cohort. Synthetic scenarios; counts do not establish physical reliability.
- nominal · placement: 4/4 completed; 4 numerically valid.
- nominal · replay_open_jaw: 0/4 completed; 4 numerically valid.
- nominal · hold: 0/4 completed; 4 numerically valid.
- low_friction · placement: 2/2 completed; 2 numerically valid.
- low_friction · replay_open_jaw: 0/2 completed; 2 numerically valid.
- low_friction · hold: 0/2 completed; 2 numerically valid.
What this advances
This environment provides a task and negative controls for a future small DVIDIA action policy. Footage review, robot actions, learned control and physical qualification remain separate stages. Rigid-block contact does not validate a soft shoe, laces or a real gripper.
Runtime and timing scope
Attempt wall time includes asset checks, compilation, reset, planning, stepping and trace capture; separate inner timings identify narrower boundaries. No rendering, sensing, training, disk serialization, transport or physical actuation.
{"platform": "macOS-26.5.2-arm64-arm-64bit", "machine": "arm64", "python": "3.12.9", "mujoco": "3.15.0", "gpu_used": false, "logical_cpu_count": 10, "process_lifetime_peak_rss_bytes": 531054592, "memory_scope": "Process lifetime high-water resident memory including imports, native models and traces, before final bundle inspection or replay export; not per-episode memory or minimum RAM. Null means unavailable."}
Each attempt retains its timing distribution, complete states, requested commands and outcome measurements. Warm model inference and camera-to-actuator latency are not measured by this authored-controller benchmark.