Event

KAI Halo Releases EgoWild2Dex Case Study: Training Dual-Arm Dexterous Manipulation from First-Person Data

1 reports · 1 sources

Kinetix AI released the KAI Halo case study EgoWild2Dex on 2026-10-23, demonstrating how to train dual-arm dexterous manipulation capabilities from egocentric data; the project page is mmlab.hk/egowild2dex. According to Kinetix AI, the method uses 538.9 hours of real, unscripted first-person data, in which backgrounds are cluttered, foreground instances are dense, and viewpoints are highly dynamic; it performs lightweight geometric alignment between the first-person view and the robot view via GeoFormer, and adopts a three-stage progressive human–robot training approach, and is said to achieve a success rate exceeding 90% on long-horizon dual-arm dexterous manipulation tasks.