Research paper
Learning from failed jumping-robot experiments
A Gaussian-process constraint model combines failure/success events with measurements available only after successful trials. Constrained Bayesian tuning is demonstrated on simulations and a physical jumping quadruped with an initially unknown constraint threshold.
Reviewed 2026-11-05 · Source and bibliography checked
- Original title
- Robot Learning With Crash Constraints
- Publisher
- Institute of Electrical and Electronics Engineers (IEEE)
- Authors
- Alonso Marco, Dominik Baumann, Majid Khadiv, Philipp Hennig, Ludovic Righetti, Sebastian Trimpe
- DOI
- 10.1109/lra.2021.3057055
- Journal issue date
- 2021-04
- Source language
- en
- Source country
- US
- Edition
- IEEE Robotics and Automation Letters
- Access
- Public bibliography and abstract
- Publication status
- Published
Source overview
A Gaussian-process constraint model combines failure/success events with measurements available only after successful trials. Constrained Bayesian tuning is demonstrated on simulations and a physical jumping quadruped with an initially unknown constraint threshold.
Use and review
Failures are allowed during learning in settings where they are undesirable but not catastrophic. This is not a method guaranteeing crash-free exploration or safe human contact.
Sources & editions
Robot Learning With Crash Constraints ↗
Institute of Electrical and Electronics Engineers (IEEE) · IEEE Robotics and Automation Letters
Archivev3abstract/authors/history, explicit IEEE DOI and RA-L6(2),1439–1446 reference checked. Failure tolerance is explicit; not described as preventing every crash.