Does an insertion policy tuned in one simulator survive a better one?

research/robofab-insert

Peg-in-hole search policies tuned by CEM in MuJoCo at three fidelity levels or with domain randomisation: tuned on the simplest level 0.91 success there but 0.53 / 0.26 on a stiffer-contact / held-out level; training on the stiffer level or with randomisation gives 0.43–0.46 held-out and cannot be ranked with 5 seeds. Stylised, sim-to-sim.

6 tests

What it found

Cross-entropy-tuned five-parameter peg-in-hole search policies in MuJoCo, trained on three simulator fidelity levels or with domain randomisation and tested on all three plus a held-out parameter set. Tuned on the simplest level, success is 0.91 there but 0.53 on the stiffer-contact level and 0.26 on the held-out one; training on the stiffer level or with randomisation gives 0.43–0.46 on the held-out set, and with 5 seeds the two cannot be ranked. No seed exceeds 0.61 on the held-out level. Toy planar task; sim-to-sim only.

A figure from its own output

Figure produced by research/robofab-insert: transfer — robofab-insert/experiments/transfer.svg, written by the project's own experiment code.

Paper, code and results

More in Robot manufacturing loops

How much of the gap between “the twin passes” and “it fails for real” does each twin fidelity close?Session-written pick-and-place programs checked in a kinematic SCARA twin at three fidelities against a held-out reference: of 69 reference-failing injected-bug mutants the levels reject 38, 46 and 60 with no false alarms; misses are unmodelled effects (stylised).
Which resource limits a small robot-printer-inspection cell, and does the busiest one tell you what to upgrade?Discrete-event twin of an arm, printers and an inspection station: printers have the highest utilisation at every size, yet at 6 printers an extra arm saves 10.9 min vs 2.3 for an extra inspection station; dispatch rules trade makespan for flow time; a mean-value twin is optimistic. Stylised, illustrative parameters.
Where is the bottleneck of an arm, printer and inspection cell?Discrete-event twin of an arm, 3D printers and an inspection station: a utilisation bound names the bottleneck (printers to 10, the arm from 12, ~12.7 good parts/h) and the simulation sits 10–11% below it near the switch; a 25% faster arm lifts 11.34 to 12.68 parts/h at 12 printers, a second inspector nothing (stylised).
Does exact queueing analysis predict a robot-printer cell's throughput better than a utilisation bound?Exact mean value analysis of the arm, printers and inspection network against a discrete-event twin: a few per cent off for exponential times, up to 11% low for less variable ones, better than the utilisation bound in most real-semantics rows, and it overshoots saturation at high work in process. Stylised.
Where does exact queueing analysis of a robot-printer cell stop working?Coefficient-of-variation sweep and a measured printer stall share for the arm-printer-inspection queueing model: MVA's error crosses zero near cv 1 while the utilisation bound stays high, and held-printer capacity times (1 − stall) reproduces the saturated rate at 4 and 8 printers. Stylised.
Is a print-failure camera worth it in a robot-arm 3D-printer cell?Is a print-failure camera worth it in a robot-arm printer cell? Blind printing against abort-on-failure: +6 to 10% good parts per hour at 2 to 10 printers, none once the arm limits the cell; break-even false-stop rate. 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.