In plain words
A navigation robot can be safe right now but still get stuck in a situation where it cannot avoid a collision a moment later.
This paper introduces SAFTER, a method that checks whether the robot could follow a safe backup route before allowing a command, then uses that same check while the navigation policy learns.
In simulated tests, SAFTER reached a training success target with fewer environment steps than ordinary SAC training. Its advantage over another safety method appeared on one shifted map, but not on the most challenging shift.
The research
A robot may meet an immediate safety limit and still lose the ability to avoid a later collision when its commands are bounded.
It remains unclear whether navigation can be learned in a way that preserves a safe recovery maneuver, rather than checking only whether each command is safe at the moment it is issued.
SAFTER filters quadruped commands using sampled recovery checks and lets the navigation policy learn through that same safety filter.
The paper asks whether a navigation policy can stay recoverable, rather than merely satisfy an immediate safety check. SAFTER answers with a filter that tests a backup maneuver and lets policy learning account for the commands the filter actually executes.
On shifted simulated maps, SAFTER outscored a matched backup-filter method on Medium-Shift, but fell below it on High-Shift. The comparison supports a benefit in one tested shift, not a uniform advantage as geometry changes.
These results cover the tested training maps, three simulated environments, and two real-world maps.
The formal guarantee is for a reduced continuous-time safety model; the robot filter runs at 10 Hz and checks backup safety at selected future times, so the theorem does not certify the full robot between updates or account for unseen obstacles.
Analysis
Deep
- Original paper authors
- Benedictus C. G. Cinun, Bin Hu
- Submitted to arXiv
- License
- CC BY 4.0