Systematic self-improvement boosts robot task success in simulation to 95%
A framework lets robots improve skills in simulation and verify them on a real robot after calibration, while keeping expectations grounded.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Yen-Jen Wang, Haozhe Jiang, Shuying Deng, Haoru Xue, Weirui Ye, Rocky Duan, Nika Haghtalab, S. Shankar Sastry, Pieter Abbeel, Haozhi Qi
- Universities and research institutions
- Not yet supplied in verified metadata; the Brief does not guess.
What they did and found
A system uses off-line data, practice in computer simulations, and analysis of failures to improve robot skills automatically. It raises the chance of completing simulated tasks from about 28.6% to 95% across 22 tasks, and then succeeds on three real tasks in 30 trials.
Why it matters
Automation can reduce manual programming and speed up safer robot use, but gains depend on the data, task variety, and the hardware used.
The system builds a library of reusable skills from offline data and tests changes across many tasks, keeping what helps overall performance.
In practice, the same calibrated system completed all three physical tasks in tests, showing some simulation gains can transfer to real work, though results vary.
What remains uncertain
Deformable objects (like towels) remain hard to handle; not all tasks improve, and real-world transfer depends on task type and hardware.
Original sources · 1
Check the original paper for its authors, methods, version and access terms.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.