7h ago

Hugging Face's LeRobot integrates TOPReward to extract zero-shot robot rewards directly from frozen video vision-language models

It evaluates task completion directly from frozen model logits.

——0——
Original post
Chris Paxton#732@CHRIS_J_PAXTONOPLeRobotLeRobot@LEROBOTHF

🤖 Zero-shot robot rewards are now in LeRobot with TOPReward! TOPReward, from @allen_ai, @UW, and @cole__ai @Amazon , turns a frozen video VLM into a robot reward model by reading log P("True" | video + instruction) directly from the model’s logits. Project: https://topreward.github.io/webpage/ Paper: https://arxiv.org/abs/2602.19313

9:29 AM · May 27, 2026 View on X
Reposted by
Ranjay Krishna#1016@RANJAYKRISHNA
QUOTE POSTChris Paxton#732Chris Paxton@CHRIS_J_PAXTON

Top reward is cool, covered on RoboPapers here: https://robopapers.substack.com/p/ep75-topreward-token-probabilities

LeRobotLeRobot@LeRobotHF

🤖 Zero-shot robot rewards are now in LeRobot with TOPReward! TOPReward, from @allen_ai, @UW, and @cole__ai @Amazon , turns a frozen video VLM into a robot reward model by reading log P("True" | video + instruction) directly from the model’s logits. Project: https://topreward.github.io/webpage/ Paper: https://arxiv.org/abs/2602.19313

4:29 PM · May 27, 2026 · 8.2K Views
5:03 PM · May 27, 2026 · 1.9K Views
Hugging Face's LeRobot integrates TOPReward to extract zero-shot robot rewards directly from frozen video vision-language models · Digg