LiberAI unveils Liber-0, a World Action Model for zero-shot robot manipulation

After pretraining, Liber-0 was evaluated on 60 tasks proposed by GPT-6, spanning pick-and-place (PnP), short Non-PnP, and challenging multi-step Non-PnP work, with 10 trials each for 600 trials total and no robot demonstrations collected for those targets[1]. Success rates were 63.5% on PnP, 34.0% on short Non-PnP and 10.5% on challenging multi-step Non-PnP, for an overall 36.0%[1][2]. LiberAI also shows demos such as cup and pouch placement, paper folding, tissue extraction, packing and storage, and boba tea preparation[1].

LiberSense Glove is a 0.3 mm, 131 g wearable that records precise finger motion with synchronized first-person and wrist cameras, so collection needs no robot on site[1]. LiberAI's LEO robot pairs dexterous hands, articulated wrists and waist, and chest/wrist camera views aligned with that glove setup to shrink the visual and kinematic gap between human demos and robot execution[1][3].

Glove recordings scale strongly for fine-grained work: 10 kh of glove data beat 2,000 kh of web human video on transfer after the same robot post-training[1]. The full staged Liber-0 recipe (broad visual prediction, joint world-action learning, then manipulation-focused sampling) reaches 83.6 PnP and 73.8 Non-PnP progress scores after robot post-training, above every single-source configuration LiberAI reports[1].

For multi-step tasks Liber-0 can optionally take intermediate goals from a reasoning agent such as GPT-6 Astra, while Astra alone scored only 8.3%, 5.0% and 1.7% on the three difficulty bands of the same suite and struggled with contact-rich alignment[1]. LiberAI says next steps include richer force and tactile supervision without losing the glove's lightweight design, plus a larger and more diverse glove collection across people, tasks and everyday environments[1][3].

Most robot policies still lean on task-specific teleop. Liber-0's bet is that action-grounded human experience, at million-hour scale and with embodiment-aware hardware, can unlock usable zero-shot manipulation and cut how much robot data each new skill needs[1][2].