JD
J. Dimsa
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Autonomous search and rescue (SaR) requires an agent to explore an unknown, partially observable and hazardous environment quickly without harming itself or the victims it is looking for. Reinforcement learning (RL) produces adaptive policies but offers no safety guarantee, while model-predictive control (MPC) enforces model-based safety constraints but is conservative and slow. We combine the two, shielding a learned policy under an MPC backup and ask when the result is preferable to either on a partially observable grid-world SaR task. We first build a pure-RL explorer and show its coverage hits a ceiling near 70% success, a limit that is due to the method rather than to tuning. We then wrap it in a slip-aware safety shield that certifies every action's reachable "slip tube" against the agent's belief and defers to a model-predictive backup when the policy proposes an unsafe move. Across 250 held-out hazardous maps, the shield makes pure RL deployable by eliminating its collisions in evaluation, reducing it from 117.7 to 0 and cutting damage around 15x while still retaining 93% of its success (from 0.43 to 0.40), at 53 ms per decision and beating the slower tube-based MPC on coverage, victims and speed. Only the exhaustive MPC is more effective, but it runs at approximately 5x the latency and does not scale. We thus position the shield not as a universal winner, but as the best choice when collisions cannot be tolerated and decisions must stay fast.
...
Autonomous search and rescue (SaR) requires an agent to explore an unknown, partially observable and hazardous environment quickly without harming itself or the victims it is looking for. Reinforcement learning (RL) produces adaptive policies but offers no safety guarantee, while model-predictive control (MPC) enforces model-based safety constraints but is conservative and slow. We combine the two, shielding a learned policy under an MPC backup and ask when the result is preferable to either on a partially observable grid-world SaR task. We first build a pure-RL explorer and show its coverage hits a ceiling near 70% success, a limit that is due to the method rather than to tuning. We then wrap it in a slip-aware safety shield that certifies every action's reachable "slip tube" against the agent's belief and defers to a model-predictive backup when the policy proposes an unsafe move. Across 250 held-out hazardous maps, the shield makes pure RL deployable by eliminating its collisions in evaluation, reducing it from 117.7 to 0 and cutting damage around 15x while still retaining 93% of its success (from 0.43 to 0.40), at 53 ms per decision and beating the slower tube-based MPC on coverage, victims and speed. Only the exhaustive MPC is more effective, but it runs at approximately 5x the latency and does not scale. We thus position the shield not as a universal winner, but as the best choice when collisions cannot be tolerated and decisions must stay fast.