AgiBot introduces WorldOracle benchmark for action-conditioned world models - 2026-10-13 - Research Publication - The 335-example evaluation spans AgiBot G2, YAM and human-hand embodiments with successful, failed and extended interactions. Seven world models are tested across rigid, deformable, fluid and articulated-object… - source: arxiv.org
AgiBot releases GE-Act 2.0 world-action model - 2026-09-09 - Research Publication - Scaling co-training data from 300 to 30,000 hours raised mean success from 17.1% to 44.1% on G1-OP and from 13.4% to 31.1% on G2-90D. The architecture combines a control-oriented autoencoder, single-step visual planner and… - source: agibot.com.cn
AGIBOT publishes LWD fleet-scale reinforcement learning for deployed generalist robots - 2026-04-30 - Research Publication - AGIBOT Finch published Learning while Deploying as a fleet-scale offline-to-online reinforcement learning framework for continual post-training of generalist VLA policies. The official page evaluates LWD on a fleet of… - source: finch.agibot.com
AgiBot X2-N study validates transformable locomotion - 2026-04-23 - Research Publication - Physical experiments covered wheel-legged, foot-legged and hybrid motion, including ramps, stairs, push recovery and skating-like travel. A NanoPi RK3588 board handled system control, with a 10,000 mAh quick-release… - source: arxiv.org
AgiBot releases AGIBOT WORLD 2026 dataset - 2026-04-07 - Research Publication - The G2-based collection combines Swift Picker grippers and OmniHand configurations with RGB-D, touch, lidar, IMU, joint-state and force information. Task descriptions, action steps, atomic skills and object labels include… - source: agibot.com.cn
AGIBOT publishes SOP framework for fleet-scale online VLA post-training - 2026-01-06 - Research Publication - AGIBOT Research introduced SOP as a scalable online post-training framework for Vision-Language-Action models running across robot fleets in the real world. The official writeup says SOP enables online updates across… - source: agibot.com
AGIBOT publishes UniFact embodied reasoning and action framework - 2026-01-01 - Research Publication - AGIBOT Finch published UniFact, presenting GenieReasoner and the FACT action tokenizer to unify embodied reasoning and robotic action inside a single autoregressive transformer. The official page positions the work as a… - source: finch.agibot.com
AGIBOT publishes Act2Goal goal-conditioned policy framework - 2026-01-01 - Research Publication - AGIBOT Finch published Act2Goal as a unified autoregressive framework that jointly pre-trains vision-language reasoning and robotic action generation. The official research page highlights strong simulated benchmark gains… - source: finch.agibot.com
AgiBot researchers test WholeBodyVLA on X2 - 2025-12-11 - Research Publication - The prototype X2 used seven-joint arms with OmniPicker grippers, six-joint legs, a waist joint and an egocentric RealSense D435i camera. Evaluation covered bag packing, box loading and pushing a 50 kg cart, with 50… - source: arxiv.org
Articles
Physical AI capital moves — 12–18 October 2026 - 2026-10-18 - Funding, acquisitions, strategic investments and public-market milestones across robotics, autonomous systems and the technologies that enable them.