cs.RONov 7, 2026

Deploying Physically and Epistemically Compliant Agents by Distilling Belief-Space Task and Motion Planners

Authors: Phillip J. K. Christoffersen, Rachel Ma, Patrick Barragán, Dylan Hadfield-Menell

Organizations: MIT CSAIL.

Abstract

In this work, we train agents to act under uncertainty according to specified, belief-dependent rules. Deep reinforcement learning (RL) can in theory train agents to follow rules if encoded into well-designed rewards, but in practice, RL fails to perfectly optimize these rewards, leading to ineffective or noncompliant agent behavior. Belief-space planners by construction produce behavior in line with developer specification, but require hand-programming online planners and belief-state estimators, adding significant engineering complexity at scale. We combine the two: we plan offline to generate a dataset of sample trajectories and then distill the plans into a policy that acts on raw observations. We propose two changes to planner distillation that address problems specific to history-based policies: Dual Averaging, a loss that combines per-timestep predictive loss and average loss on task changes, and ρρ-Optimal Reinforcement, a feedback mechanism that reinforces the policy's own successful rollouts when these are near-optimal. We provide a sufficient condition under which static search problems with noiseless observation models is complete, and show that maximum-likelihood-observation planning 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% in 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.

Figures & tables

Explore similar work

May 25, 2026

Learning to Search and Searching to Learn for Generalization in Planning

Combinatorial generalization remains a central challenge in Deep Reinforcement Learning (DRL). Classical planning provides a simple yet challenging setting to study this problem through explicit relational descriptions, without requiring learning from perception. In sparse-reward domains, standard RL exploration via real-time search is ineffective, and learning-based planning methods often rely on expert demonstrations, hindsight relabeling, or random walks from the goal state. In contrast, planners rely on best-first search methods such as A⋆\mathrm{A}^\star to solve problems from scratch. We propose a self-improving WA⋆\mathrm{WA}^\star learning framework in combination with a value heuristic represented by a Relational Graph Neural Network: the heuristic guides search, and the resulting search data updates the heuristic via QQ-learning. This loop yields heuristics that can function as general policies and solve new instances even without search, where DRL otherwise fails, as we show on puzzles such as Sokoban, PushWorld, The Witness, and the 2023 International Planning Competition benchmarks. Notably, we demonstrate strong zero-shot generalization: For example, heuristics trained on Blocksworld instances with fewer than 30 blocks successfully solve instances with 488 blocks without search.
Sep 29, 2026

State Trace Rationale As Auxiliary Task in Reinforcement Learning

We propose STRAT, an auxiliary task that trains deep reinforcement learning (RL) agents to predict a short textual trace of their own state. Inspired by human spatial navigation, the description combines landmark, route, and survey knowledge, tracking the agent's position, inventory, goals, and immediate progress. Environment rules generate this text online without human labelling. Our method adds a single auxiliary head to a standard policy. Across 60 sparse-reward XLand-MiniGrid tasks, STRAT solves complex environments where standard RL fails outright, while compacting state representations and preventing rank collapse. Beyond performance gains, the predicted trace provides a readable account of agent beliefs at every step for no extra cost.
Nov 12, 2026

What Can a VLA Learn from the Agent That Calls It? Task-Level Experience Distillation for Embodied Navigation

Vision-language-action (VLA) models drive the robot directly from camera views, yet a VLA alone shows three typical failures. It walks past the goal, stops before reaching it, and rarely recovers after going off track. Pairing a VLA with a large language model (LLM) planner fixes these failures, but queries the LLM throughout every episode, adding token costs and latency. We introduce Harness-to-Executor (H2E), a task-level distillation method that pairs an LLM planner and a frozen VLA as a heterogeneous training agent, keeps only sessions that reach the goal, and then trains a VLA alone on them so no LLM is used at test time. In these sessions, the planner gives sub-instructions and moves the robot with its own tools, detours are common, and the planner decides when to stop. H2E therefore labels every motion as a VLA action for the original task, whoever made it. It masks the loss on steps moving away from the goal and assigns end labels to states near it, so the VLA learns where the task ends. H2E sets a new state of the art on REVERIE-CE and NavRAG-CE. It raises an 8B VLA's success rate by 15.6% and 11.1% respectively, surpassing a method trained on 10M samples. The distilled VLA exceeds its agent on unseen episodes with zero LLM tokens and runs 3.7 times faster.