UniPred, accepted at IEEE T-RO 2026, invents task concepts from demonstrations for long-horizon robot planning
tomssilver · x · 2026-10-28
UniPred, a paper from CMU, UMich, Princeton and partners, is accepted to IEEE Transactions on Robotics 2026 (SI on Foundation Models for Robotics). The core problem: robot planners rely on predicate concepts over long-horizon tasks, but demonstrations only label actions. UniPred invents those concepts with an agent-in-the-loop scheme: an LLM agent proposes when a concept becomes true/false, a neural classifier learns it from demonstration image features, and inconsistent proposals stay high-loss, prompting the agent to revise until concepts separate cleanly. A planner chains learned concepts into long tasks with recovery when a skill misfires; the same learning works from object states, point clouds, and camera images. Authors include Tom Silver, Katia Sycara, and Sebastian Scherer.
More from Embodied
- Foundry Robotics raises $100M Series A to build 130,000 sq ft factory in Emeryville, California — ishabytes · 2026-10-29
- Sony launches ELF-MN1, a 6-inch glasses-free spatial display that puts VRChat avatars on your desk — TinfoilTricorn · 2026-10-29
- NVIDIA's Metatron lets coding agents drive humanoid whole-body motion via a goal interface — carlosdponx · 2026-10-29
- StackChan Desk Robot Runs Grok + Hermes Agent, Lights Sync With Mood — alexcovo_eth · 2026-10-29
- Purdue's GPS-free drone tracks runners through forests; Rachel Thomas asks if it should exist — mmitchell_ai · 2026-10-29
- Insta360 X6 Spatial Capture Turns a 360° Walk Into an Explorable Gaussian Splat Scene — Aiden_Tech_Ai · 2026-10-28