Johns Hopkins study compares four failure data sources for robot reward models
mangahomanga · x · 2026-11-06
A Johns Hopkins paper, "The Right Kind of Wrong," asks which counterfactual failures to collect when training reward models for robot RL. It compares kinesthetic teaching, policy rollouts, human videos, and real2sim rollouts across five real manipulation tasks, evaluates via LoRA-adapted and frozen reward models on a 250-rollout human-annotated benchmark, and uses the best reward models to guide RL post-training of π₀ to test whether offline reward gains translate to better real robot behavior.
Related event: JHU Study Compares Four Failure Data Sources for Robot Reward Models(7 posts)→
More from Embodied
- Modded MacBook Pro gets a working touchscreen — used to play Honor of Kings — xiaohu · 2026-11-06
- Roboticist shares Stanford-collab research directions and BARS talk — Fowe · 2026-11-06
- RAX From Harvard/NVIDIA/Berkeley Lifts Robot Success From 26% to 94% With Frozen Weights — jiqizhixin · 2026-11-06
- Open-Weights Robot-Use Harness: Qwen3.8 Scores 31.39, Beating GPT-6 Astra's 28.97 — aigclink · 2026-11-06
- Dev builds Salvo battleship game for dual-screen iPhone Duo with head-to-head and AI play — HankYeomans · 2026-11-06
- First-hand demo of Wayve AI Driver on San Francisco streets impresses rider — Sethwinterroth · 2026-11-06