Back home

Liber-0

x+web 0.90
Visit source

Intro

Liber-0 is a World Action Model (WAM) from Beijing physical-AI startup LiberAI (technical post). It is pretrained on 2.2M+ hours of human experience, combining action grounding from the LiberSense Glove, a Unified Embodiment Representation, and the human-aligned LEO robot platform to transfer everyday manipulation skill into robot policies.

On 60 GPT-6–proposed tasks / 600 trials, Liber-0 reports 36.0% overall zero-shot success (63.5% PnP / 34.0% short Non-PnP / 10.5% multi-step), without robot demos collected for those target tasks. Company posts plus Access Newswire, Robotics & Automation News, and Chinese tech coverage circulated around 2026-11-03–11-04; earlier radar passes had not stored the name—this scan inserts on blog + press + X corroboration.

Explain

Problem

Classic robot manipulation depends on costly, task- and robot-specific demonstrations. Liber-0 aims to scale from human experience—video plus measured glove motion and embodiment alignment—so robots can attempt new tasks with far fewer target-task robot demos, including some multi-step skills.

Differentiation

  • Framed as a World Action Model that jointly predicts future frames and actions.
  • LiberSense Glove supplies scalable action anchors without an on-site robot; LEO reduces visual/kinematic shift vs human demos.
  • Publishes stratified zero-shot rates on a GPT-6–authored task suite.

Limits

  • Multi-step success remains modest (~10.5%); treat claims as research/product preview until independently reproduced.
  • Weights, API, and commercial access follow LiberAI announcements.
  • Pairing with planner agents (e.g. Astra) is a workflow story, not a guaranteed SLA.

Usage

Who it is for

Embodied-AI researchers and robot-product teams watching zero-shot manipulation and human-to-robot skill transfer; evaluators tracking WAM / glove-collection recipes.

Getting started

  1. Read the Liber-0 technical post and homepage demos; note the exact task definitions and success metrics.
  2. Watch LiberAI channels for weights, interactive sim demos, or LEO/glove partnership interfaces.
  3. For reproduction, keep the PnP / Non-PnP / multi-step split rather than cherry-picking highlight clips.

Advertisement