LiberAI Introduces Liber-0 Model for Zero-Shot Manipulation in Robots

Summary

LiberAI released Liber-0, a World Action Model that learns robot manipulation from human hand movements recorded with a 131-gram sensor glove. The model achieved an overall 36% success rate on the 60 zero-shot tasks proposed by GPT-6. The company claims that 10,000 hours of glove data outperformed 2 million hours of internet video under matching conditions.

Why it matters

This approach reimagines robot skill acquisition around passive, low-cost human motion capture rather than expensive teleoperation, potentially significantly lowering barriers to deploying agile robots. The claim that glove hours surpass web video challenges the prevailing 'more video is better' assumption in embodied AI. However, the benchmark is entirely self-administered, and independent replication is still needed for these figures to carry real weight.

Highlights

• Overall zero-shot success rate of 36%; only 10.5% on multi-step tasks
• The 131-gram LiberSense Glove records hand poses and trajectories without a robot in the loop
• 10,000 hours of glove data outperformed 2 million hours of internet video; 25% robot demonstrations + glove data scored 75.6, while full demonstrations scored 72.8

Critical take

The 'zero-shot' label is generous: the model was still pre-trained on over 2.2 million hours of human data and some robot demonstrations. The 60-task benchmark was both proposed and evaluated by the same team, with no third-party verification. The 10.5% success rate on multi-step tasks is far from deployable levels. The self-reported 56/100 AI score in the source post also indicates that the community's initial reaction has been lukewarm.