
Robot Arms
SO-100
TOPReward official materials report ManiRewardBench evaluation across SO-100 and SO-101 platforms.
Applies to the reported ManiRewardBench evaluation protocol.
Participating organizationsUniversity of WashingtonParticipating organizationsHugging FaceParticipating organizationsAllen Institute for AI
Participating organizationsAmazonParticipating organizationsUniversity of North Carolina at Chapel HillVersion scope and roles for these associations await verification.
TOPReward extracts dense task-progress rewards from pretrained video-language model token probabilities without task-specific reward training. It supports success detection and reward-weighted imitation learning across real robot manipulation tasks.
The original paper introduces training-free token-probability rewards for robot trajectory evaluation.
This announcement
The original paper introduces training-free token-probability rewards for robot trajectory evaluation.
First arXiv disclosure of the original method, not a verified launch date of the later LeRobot implementation or a new model-weight release.
Paper available
Author order: Original order
Original method paper referenced by the LeRobot implementation; not a separate evaluation of every LeRobot integration change.
Hugging Face LeRobot · Documentation
Publication date unknown · Accessed: 2026-07-24
TOPReward authors · Source
Publication date unknown · Accessed: 2026-11-07
Title not provided
Publisher not provided · Type unconfirmed
Publication date unknown · Accessed: 2026-10-19
TOPReward authors · Code repository
Publication date unknown · Accessed: 2026-11-07