Amap and Alibaba Token Hub Release Humanoid Robot Model λ₀, Trained on 500 Hours of Human Video Data
Amap and Alibaba's Token Hub business group have jointly launched the embodied intelligence model λ₀, attempting to train humanoid robots to perform full-body coordinated tasks such as walking, carrying, and two-handed manipulation using videos of everyday human activities and motion data. The team built the HumanVerse dataset, which contains 500 hours of human activity data in open environments, collecting first-person videos via wearable devices while synchronously recording body motion and reconstructing hand movements, then converting them into robot training data through viewpoint matching and motion alignment. λ₀ adopts a three-stage training pipeline and has been deployed on the Unitree G1 humanoid robot; the project's official website shows that it achieves a 90.0% success rate on the SIMPLE benchmark and completed 25 of 40 trials across four real-world mobile manipulation tasks, a 62.5% success rate, 22.5 percentage points higher than the strongest baseline compared in the paper.
Why read this
AI score 58
The original text presents λ₀'s three-stage training pipeline and results on two benchmarks, offering insight into the specific approach of training humanoid robots with human video data.
Topics
Body · AI translation
On October 26, Amap and Alibaba's Token Hub business group jointly launched the embodied intelligence model λ₀, attempting to use videos of people's daily activities and motion data to train humanoid robots to learn tasks requiring whole-body coordination, such as walking, carrying, and two-handed manipulation.
According to the project team, the research constructed the HumanVerse dataset, containing 500 hours of human activity data collected in open environments. The data was recorded via wearable devices as first-person videos, with body motion information collected synchronously and hand movements reconstructed; after viewpoint matching and motion alignment, it was converted into training data suitable for humanoid robots.
In terms of training, λ₀ adopts a three-stage training pipeline: first learning object interaction and hand movements from diverse first-person human videos, then using HumanVerse data to train coordinated control of the body and hands, and finally adapting with data targeted at robot tasks. The team applied the relevant methods to the Unitree G1 humanoid robot, exploring the transfer of human motion experience to robot control capabilities.
According to experimental results published on the project's official website, λ₀ achieved a success rate of 90.0% on the SIMPLE benchmark; in four real-world mobile manipulation tasks, the model completed 25 of 40 trials, for a success rate of 62.5%, 22.5 percentage points higher than the strongest baseline compared in the paper.
Currently, robot training still faces problems such as the high cost of collecting high-quality demonstration data and limited task and scenario coverage. λ₀ attempts to introduce human whole-body motion data, which requires no robot participation to collect, into the training pipeline, offering a new approach for expanding the scale of embodied intelligence training data.
Submissions/leaks: tougao@youxituoluo.com
Article/business cooperation: Lin Nan (WeChat 19250561593)
Join the industry exchange group: Lin Nan (WeChat 19250561593)
Metaverse Digital Industry Service Platform
Download the "Tuoluo Technology" APP to get cutting-edge in-depth metaverse information