ACQUIRING TARGETS...
· DroidHunt news
LiberAI (将闲科技), the Beijing world-model startup, unveiled Liber-0, a "World Action Model" pretrained on more than 2.2 million hours of human experience, per The Robot Report (Nov 3) and the company blog. The full release follows September's Liber-0 Preview, which topped the RoboDojo benchmark. The recipe: three-stage pretraining (visual prediction, joint world-action prediction anchored by LiberSense glove recordings, manipulation focus) plus an embodiment representation for cross-body transfer. On 600 zero-shot trials of GPT-6-proposed tasks — no task-specific robot training — the company reports 63.5% success on pick-and-place, 34.0% on short non-PnP and 10.5% on multi-step tasks (36.0% overall), and claims glove data alone beats 200× more web video. All figures company-reported.
Why it matters: Action-grounded pretraining from cheap sensor-glove data is a direct attack on the robot-data bottleneck: if glove-scale supervision transfers this well, the economics of teleoperation-heavy data collection shift. Independent verification is still pending.
Sources: therobotreport.com liberai.com
Researchers from BIGAI, USTC, SJTU and PKU present SoccerBot, a hierarchical reinforcement-learning framework whose latent motion prior is distilled from human soccer motion-capture data (~2.5 hours of curated references after feasibility filtering). A single skill-conditioned policy learns moving, dribbling, trapping and kicking — claimed as the first real-world humanoid soccer system combining the full repertoire with human-like motion and fluid transitions. Trained in NVIDIA Isaac Lab and deployed zero-shot on a Unitree G1: simulated success rates of 97.0% (kicking), 89.9% (dribbling) and 100% (moving); real-world kicking-direction success of 95%/85%/70%/80% across four ball conditions, a ~3.1 percentage-point average drop vs simulation. Real tests relied on an external motion-capture system for ball state. Preprint.
Why it matters: Sports are a proving ground for the whole-body control and contact-rich skills that commercial manipulation needs. The gaps it names — mocap dependence, rigid-ball contact models — are exactly what blocks outdoor, uninstrumented deployment.
Sources: arxiv.org
XROGO, a Chinese robot "work network" operator, used CeMAT ASIA 2026 (Nov 3) to launch its "robot on-post plan" for enterprise customers and logistics partners, per a company-provided QbitAI article. Initial roles: machine tending (machine-tool and heat-treatment loading, pick-and-place around equipment) and inspection/duty watch in server rooms and power-distribution rooms, with warehouse sorting, loading and tallying to follow. The model is "hire robots like people": rental-plus-service contracts around agreed tasks, frontline workers and experts as robot instructors, and remote staff authorizing actions and handling exceptions. An accompanying recruitment post offers up to four months of free equipment rental, limited to 50 units.
Why it matters: RaaS-plus-teach-in bundles are how operators are trying to lower deployment friction, and the 50-unit, four-month-free offer is a concrete price signal for real-job pilots — watch conversion to paid contracts.
Sources: qbitai.com
Tags: xrogo deployment china company-claimed
Home Every digest Robot catalog Drones on DronePriceHunt Privacy & affiliate disclosure RSS