Amap open-sources Lambda-0 whole-body humanoid model

Amap, Alibaba's mapping unit, released Lambda-0, an open-source whole-body foundation model for humanoids trained solely on human first-person data collected without robots. Code is on GitHub under Apache 2.0 with the 500-hour HumanVerse-500 dataset. Zero-shot transfer to a Unitree G1 reached a 62.5% real-robot success rate, 22.5 points above the best baseline.

• One demo takes ~29 seconds to collect vs ~100 seconds with a robot
• HumanVerse-500 dataset holds 500 hours of whole-body human data
• Zero-shot transfer to Unitree G1 using only human data
• 62.5% real-robot success, 22.5 points above the strongest baseline