A different question from brief proximity
The previous RoboLLM repair study reached a moving object’s center within 3 cm in 35/100 scenes. Six of those successful episodes also moved the target by more than 3 cm. That criterion measured brief proximity; it did not require a stable approach.
This study asks whether the arm can approach a fixed point above an object, hold there and leave the scene substantially unchanged. It introduces a different task. Its score cannot be compared directly with the earlier 35/100, and the old trajectories are not re-scored.
Define the goal, holding interval and failures together
The goal keeps the requested object’s initial X/Y coordinates. Its height is 10 cm above the highest point of that object’s compiled initial collision geometry. It stays fixed even if the object moves. This is a standoff task: the hand approaches an empty point above the object, without an orientation requirement or grasp.
- Reach and hold: the end-effector reference must remain strictly within 2 cm for one second: 501 qualifying samples spanning 500 intervals of 2 ms. Leaving the radius resets the holding interval.
- Preserve every object: a conservative bound on surface displacement must stay at or below 5 mm throughout the episode. It combines body translation with a rotation term, using the object’s enclosing radius. Rotation therefore counts even when its center hardly moves.
- Avoid forbidden contacts: any generated contact at zero or negative signed distance involving a moving robot body fails the episode, including self-contact and contact with the base.
Object/table, object/object and static mounting contacts are allowed, subject to the motion bound. Checks run from reset at 500 Hz; violations take precedence over success. There is no settling phase. The ten-second horizon includes holding, and starting inside the radius still requires a full second. No failed scene is removed.
Two controllers, identical physics and starting states
Both controllers use the repaired simulator, ideal robot gravity compensation, open fingers and the same fixed goal. Complete initial integration states, model parameters and object configurations must match within each pair. Joint commands arrive at 20 Hz; the independent execution monitor checks every 2 ms.
Direct DLS retains the gain-8 Cartesian command and the existing damped Jacobian mapping to joint targets. The collision-aware joint planner searches up to 64 inverse-kinematics starts, with 250 iterations each. It screens endpoints and straight joint-space paths on a copied model with 3 mm robot-geometry margins, then executes the shortest accepted joint displacement with a quintic position profile. It also checks the connector from the observed joints to each next command.
Planning samples paths at at most 0.02 radians per joint. An unsuccessful search or rejected connector counts as failure; neither proves that the goal is unreachable. Planning time does not advance the simulation, and CPU durations are recorded separately. This is not an equal-compute or real-time comparison.
Results and inspectable evidence
The frozen protocol reserves seeds 260928000–260928099 for 100 paired scenes. Development used 17, 18 and 101–110; both earlier evaluation ranges are excluded.
| Controller | Successes | Wilson 95% interval |
|---|---|---|
| Direct DLS | 29 / 100 | 21.01–38.54% |
| Collision-aware joint plan | 88 / 100 | 80.19–93.00% |
Both controllers succeed on 29 scenes; the joint planner alone succeeds on 59, and both fail on 12. The observed difference is +59 percentage points. Direct DLS fails on 71 forbidden-contact episodes. The planner’s 12 failures are unsuccessful finite path searches, not proven unreachable goals.
The 200 episodes contain 252,950 physics intervals. The largest all-object surface-motion bound remains below 0.118 mm across observed samples, against a 5 mm limit. Observation stops at success or failure; durations differ between episodes.
Recorded trajectories and paired outcomes support the aggregate score. Per-controller Wilson intervals do not describe uncertainty in the paired difference.
An independent check of all 253,150 recorded samples confirmed the full holding interval for all 117 successful episodes. A second execution matched the complete semantic record and provenance. This verifies repeatability on that runtime, not additional independent scenes.
Sampled constraints are not a safety guarantee
Both controllers use privileged simulator state and a known model. This tests neither visual grounding nor a learned policy, grasp, placement or physical robot. Ideal gravity support remains a simulation assumption.
MuJoCo’s unchanged collision masks and body-pair filters omit some pairs, including adjacent bodies. Planning and execution checks are discrete, not continuous geometric certification. A passing score establishes only the declared position, motion and generated-contact conditions over the recorded samples.
Reproduce the contract as well as the score
The method and reproduction guide fixes the runtime, commands and parameters. The runner requires a clean frozen revision for reserved scenes and retains aborted partial evidence on unexpected execution errors. The planner implementation records its work counters.
The frozen source passed 293 focused tests locally and in GitHub validation, alongside three development simulation smokes. These checks do not certify the full training/rendering suite. Implementation and regression tests were developed with Codex assistance.