Deploying Physically and Epistemically Compliant Agents by Distilling Belief-Space Task and Motion Planners
Deploying Physically and Epistemically Compliant Agents by Distilling Belief-Space Task and Motion Planners
This paper presents a method that combines offline belief-space planning with policy distillation to create agents that act under uncertainty according to specified, belief-dependent rules. The authors address challenges in history-based policies through Dual Averaging loss and ρ-Optimal Reinforcement feedback. They introduce a new domain, Scavenger, where agents must gather information to succeed, evaluating the distilled policy on both visual and natural language command inputs. Results show the learned policy achieves over 90% success in Scavenger versus at most 5% for standard reinforcement learning, with strong transfer capabilities to unseen commands and larger floorplans.
The authors combine offline belief-space planning with policy distillation to address reward hacking in deep reinforcement learning, achieving over 90% success in a new information-gathering domain called Scavenger compared to at most 5% for standard RL.
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
Reported success rate
90%
View original evidence
lanning implies this condition in static search with continuous beliefs. We create a challenging new domain, Scavenger, where an agent needs to gain information in order to succeed. We evaluate our distilled policy in two versions of Scavenger, one where information is provided visually and one where it is provided as a command in natural language. In either case, the learned policy attains a success rate above 90% i
Open source S5
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
lanning implies this condition in static search with continuous beliefs. We create a challenging new domain, Scavenger, where an agent needs to gain information in order to succeed. We evaluate our distilled policy in two versions of Scavenger, one where information is provided visually and one where it is provided as a command in natural language. In either case, the learned policy attains a success rate above 90% i
Open source S5
n Scavenger, compared with at most 5% for RL. We also find that distilled planners transfer with greater than 60% success to following unseen and intentionally vague commands and greater than 75% success to following floorplans larger than those seen in training.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org