Sequential Recovery Updates for Learned Controllers under Successful-Trajectory Preservation
Sequential Recovery Updates for Learned Controllers under Successful-Trajectory Preservation
This research introduces a framework for improving learned controllers in humanoid robots by learning correction mechanisms without compromising existing successful behaviors. The method constrains entry rules rather than skill parameters, ensuring that updates only occur when the composed controller can reproduce new recovery strategies while exactly replaying all protected trajectories. Validated on two humanoid systems, the approach achieved 62 additional recoveries over 192 withheld perturbations without losing any of the base controller's successes. The study further analyzes the trade-offs involved in relaxing these constraints, demonstrating that removing them leads to failures in completing previously protected tasks.
The study presents a method for updating learned controllers in humanoid systems while strictly preserving previously successful trajectories, validated on two distinct platforms with reported recovery improvements under specific perturbation conditions.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
recoveries
62
Reported trials
View original evidence
m runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.
Open source S4perturbations
>192
Reported trials
View original evidence
m runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.
Open source S4conditions
33
Reported trials
View original evidence
m runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.
Open source S4protected completions lost
8
Reported trials
View original evidence
m runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.
Open source S4recoveries gained by removing constraint
33
Reported trials
View original evidence
m runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
m runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org