Do the robofab carry results survive an independent physics engine?

research/robofab-mujocotorque

Re-checks the 108 trapezoid and time-optimal carry profiles of the robofab-torquebound rows against MuJoCo inverse dynamics: the hand-written two-link torques agree to 1.8e-5 N m, peaks are within 0.73% of the 5 N m limit (stylised, joint torques only).

5 tests

What it found

Follow-up to robofab-torquebound (stylised, preliminary). The two-link arm behind every robofab carry result is rebuilt in MuJoCo and the 108 actual profiles (best trapezoid and time-optimal, nine rows, six parts each) are checked with MuJoCo inverse dynamics. The hand-written joint torques agree to at most 1.8e-5 N m against a 5 N m limit, so the 9 to 23% edge is not an artefact of a dynamics bug. Where only torque binds, MuJoCo puts the trapezoid's peak at up to 5.037 N m (0.73% over) and the time-optimal profile's at up to 5.011 N m (0.21% over). An open-loop replay drifts up to 0.023 rad and 1.4 mm. Joint torques only: the jaw-force bound and motor model are not checked. Planar toy arm, assumed parameters; simulated only.

A figure from its own output

Figure produced by research/robofab-mujocotorque: figure — robofab-mujocotorque/experiments/figure.png, written by the project's own experiment code.

Paper, code and results

More in Robot manufacturing loops

Does the time-optimal carry's edge over a tuned trapezoid survive when torque or the jaw cap binds?Re-checks the torque- and jaw-bound rows of robofab-scurve (no speed limit) with a converged solver and denser trapezoid ramp sampling: the time-optimal carry is still 9.4 to 23.0% faster than the tuned trapezoid in all 9 rows, and the edge is stable to solver node count (under 0.1%) and rises by at most 0.4 points with denser sampling. Stylised.
Once the time-optimal solver's node count is converged, how big is its edge over a tuned trapezoid?Re-solves the time-optimal carry of robofab-scurvefine at 400 to 25600 nodes in all 12 rows: converged to 0.02% by 6400 nodes, the 400-node answer was up to 0.5% pessimistic, and the converged edge over a ramp-resolved trapezoid is 0.11 to 0.88% in every row; stylised, preliminary.
How hard must the jaws squeeze while the arm moves?How hard must the jaws squeeze while the arm moves? Simulated minimum force matches the m·sqrt((g+a)²)/(2μ) bound within 7% over 77 cases and is at most 0.3 N at 40 m/s²; a contact time constant that passes the static test reports slip at any force once the jaws accelerate (stylised).
Does the jaw-force bound still hold at arm accelerations of 100 to 800 m/s²?Does the jaw-force bound still hold at arm accelerations of 100 to 800 m/s²? With fine MuJoCo contact settings the minimum force is 0.89–1.02× the bound over 39 cases; the earlier settings fail above 200 m/s² (stylised, prescribed acceleration, no arm).
How fine must the MuJoCo contact settings be for a moving grasp?How fine must MuJoCo contact settings be for a moving grasp? On all seven toy-gripper parts, solref 0.2 ms and below reproduce the jaw-force bound (0.98–1.04×) at 100–800 m/s²; 0.5 ms fails at 400 and above, and the adequate value falls with acceleration (stylised, upward load only).
Is it the contact time constant or the timestep that breaks a moving grasp?Is it solref or the timestep that breaks a moving MuJoCo grasp? On a 24-cell grid only 6 cells hold; a coarse step always fails, a finer step does not rescue a marginal solref, and solref 0.1 ms holds at steps of 0.1 ms and below (stylised).

All 53 Robot manufacturing loops projects →

Other research topics

Proper scoring & elicitation28 Effort & peer prediction9 Verification games & slashing19 Detection & tolerances14 Markets as routers13 Prediction-market design9 Paying contributors8 Decentralised training systems23 Digital twins & sim-to-real164 LLM chip design loops40 Robot manufacturing loops53 Ambient AI assistant46 Self-replicating manufacturing34 Simulation-first design64 Engineered clothing45 Task-specific AI accelerators40 Algorithmic chip-design loops42 Open-source purpose-built robots44 The software-engineering loop43 Adversarial human imitation41 Open-source residential construction41 Autonomous multi-angle capture34 Sensing, imaging and 3D31 Energy supply and dense energy storage39 Robot societies among people31 Agentic co-design of low-cost robots27 Laundry-hanging robots25 Machine learning on qualitative data23 Robots on the building site27 Civilizations & group psychology1 History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.