ACE Robotics and CUHK MMLab open-source ACE-Ego embodied VLA model
ACE Robotics and CUHK MMLab released ACE-Ego, a one-brain, multi-type embodied-operation VLA model trained with egocentric human and robot data.
Evidence notes
- ACE-Ego reached 72.8% average success on RoboCasa GR1 TableTop for humanoid robots.
- It reached 90.62% on RoboTwin 2.0 domain-randomization tests for dual-arm manipulation.
- Target tasks include retail sorting, order fulfillment, packing plastic bags and boxing shoes.
Company context
ACE Robotics develops embodied-intelligence world models and robot-brain modules, including Kairos, A1 and ACE-Ego. Its June 2026 ACE-Ego release paired egocentric human and robot data for multi-type embodied-operation VLA tasks across humanoid and dual-arm manipulation benchmarks.