Princeton-led work uses coding agents to synthesize reusable policies for generalized task and motion planning
tomssilver · x · 2026-09-25
A Princeton team led by Tom Silver, with CMU, Cambridge and FBK collaborators, released AgenticGenTAMP: a study evaluating whether coding agents can synthesize a single reusable programmatic policy for generalized task and motion planning (TAMP) across 28 simulated environments.
- Setup: In TAMP, discrete choices (which object, which tool, which subgoal) couple with continuous geometric, kinematic and dynamic constraints. Within an environment, instances differ in object count, configuration and geometry but share state/action spaces and rewards — the agent must generalize.
- Restrictions: The agent gets only the task description, observation/action spaces, and goal; each synthesis run has a $20 model-usage budget and may choose its own simulator tests. No network or host filesystem access; code runs in an isolated Docker container; no environment source code, only a client API.
- States are fully observed and object-centric; perception and language understanding are out of scope. Paper, code, full results and rollout gallery are on the project site.
More from Embodied
- Perceptron releases Mk1.5, one model to fly drones and control robots — AkshatS07 · 2026-09-25
- Robots doing unexpectedly clever things: problems unsolvable six months ago now cracked — tomssilver · 2026-09-25
- ROS 2 Makoa Charts a Less-is-More Path: Cutting, Not Adding — facontidavide · 2026-09-25
- Sims and Sensors Called the Most Fertile Ground at the Infra Layer — pzakin · 2026-09-25
- Robot simulation is the cleanest anti-gaming incentive, argues Bittensor SN49 — bittingthembits · 2026-09-25
- AgenticGenTAMP: coding agents synthesize reusable robot TAMP policies across 28 envs — tomssilver · 2026-09-25