LIBERO-MAX benchmark tests robot policies against mid-execution disturbances, NeurIPS 2026
jiqizhixin · x · 2026-10-17
Existing robot robustness benchmarks alter conditions before a task starts — different cup position, lighting, camera angle — so the robot sees the changed scene from frame one. But "the cup was already there" and "the cup moved mid-reach" demand very different capabilities, and most benchmarks never test the latter.
Co-first authors Yunbei Zhang and Zijian Jin, with UT Austin's Peter Stone, Stanford's Marco Pavone, MIT postdoc Weirui Ye, and NTU's Jianfei Yang, present LIBERO-MAX (NeurIPS 2026): it evaluates robot policies under mid-execution environmental changes across four disturbance categories — geometry (target/receptacle relocation), observation (camera shift, sensor noise onset), appearance (illumination/visual theme switch, distractor burst), and path (obstacle insertion). The team benchmarked 14 current robot strategies spanning VLA models.
More from Embodied
- Surface Pro Ultra review: 32GB of its 64GB memory is hardware-reserved, a headache for local AI — draginol · 2026-10-18
- lidar_slam_ros2: an 860-star ROS 2 LiDAR SLAM toolkit for Autoware-compatible mapping — rsasaki0109 · 2026-10-18
- Reviewers split on Meta's new AI glasses: camera-free version preferred, always-on scary — thisguyknowsai · 2026-10-18
- Amazon's robot fleet tops 1 million, raising questions about its 1.5M human workers — Polymarket · 2026-10-18
- Hermes robot gets a Raspberry Pi brain, stereo cameras and mics next — natesiggard · 2026-10-17
- Indie robot builder finishes body design, still puzzling over neck DOFs — _Stocko_ · 2026-10-17