Alibaba’s Amap unveils λ0 model for humanoid whole-body skills

The Alibaba-owned navigation and location services company Amap, working with Alibaba Token Hub, has introduced λ0, a foundation model intended to help humanoid robots learn coordinated movement and manipulation from human demonstrations. The system is designed to let a robot acquire whole-body skills without every behavior being built through robot-specific teleoperation.

At its core, λ0 learns from first-person video that is synchronized with recordings of body and hand motion. That pairing gives the model a view of what a person sees while also capturing how the person moves, turns, reaches and handles objects. The approach is meant to connect perception with action in a form that can be transferred to a humanoid platform.

The team says the model can combine walking, posture adjustments and two-handed operations within longer tasks. Instead of treating locomotion and manipulation as separate problems, λ0 is trained to coordinate them as parts of a single sequence. That matters for household, warehouse or service scenarios where a robot must move to an object, position its body and then use both hands.

To support the work, the researchers built HumanVerse, a dataset containing 500 hours of people performing activities in real-world environments. The dataset is described as capturing human behavior in ordinary settings, giving the model a broad source of demonstrations for how people move and interact with objects.

A central motivation is to reduce reliance on collecting demonstrations through robot teleoperation. Teleoperation can produce high-quality robot data, but it is slow, expensive and limited by the number of robots and operators available. Human demonstrations, by contrast, can be gathered with wearable or handheld equipment and then used to train policies that generalize across tasks.

In the team’s evaluations, λ0 achieved a 62.5% success rate across real-world locomotion-and-manipulation tasks. The result indicates that the model can complete a meaningful share of compound tasks that require both moving through space and interacting with objects. At the same time, the success rate shows that whole-body humanoid control from human video remains an active research problem rather than a solved capability.

The project has released data-collection hardware configuration code and sample data for researchers. The release is intended to make it easier for other labs to reproduce the collection setup and experiment with the dataset. By sharing both the hardware configuration and sample data, the team is offering a starting point for work on human-to-robot skill transfer.

λ0 is part of a broader push to build foundation models for embodied AI, where robots learn from large-scale human data rather than from hand-coded routines. The promise is that a single model can acquire reusable skills and combine them in new sequences. The challenge is ensuring that skills learned from human video remain robust when mapped onto a machine with different sensors, joints and physical limits.

The researchers frame the work as a data-centric route to humanoid intelligence. Rather than laboriously scripting each skill, the aim is to let the model infer patterns from many human examples and then adapt those patterns to robotic embodiments. Such an approach depends heavily on data quality, synchronization and coverage of tasks.

For Amap, the project also reflects the growing overlap between navigation, mapping and robotics. Amap is known for location services and route planning, capabilities that matter for robots that must move through real environments. Alibaba Token Hub’s involvement points to the company’s interest in building the data and model infrastructure needed for embodied systems.

The λ0 announcement does not claim full autonomy or deployment at scale. Instead, it presents a research direction: use human first-person video and synchronized motion data to teach humanoids compound skills, evaluate them in real-world tasks and release tools for further study. The 62.5% success rate and the HumanVerse dataset provide a baseline for measuring progress as the model and data pipeline are developed further.

For now, the public materials focus on data collection and evaluation rather than a commercial robot product. That framing places λ0 in the research stage, with the released code and sample data meant to support reproducibility and external benchmarking. The project is presented under the Alibaba λ0 project label.