Astribot's Plan to Solve Robot Generalization

In this talk

1. 0:12

   Teleoperation and low latency

   Huang says Astribot develops general models for robot manipulation. The host tries a teleoperation demo and notes its low latency; Huang attributes that to robot design and control methods optimized for remote operation.

2. 2:11

   Generalization, data, and dexterity

   Huang describes safe collaboration with people and broadly capable models as major challenges. Her central research interest is generalization, which she says depends on the right data, architecture, and training recipe. Astribot is also exploring continual learning, active gaze data, and ways to adapt human manipulation data for dexterous hands.

3. 4:54

   Pretraining and the path to home robots

   Huang says teleoperation data is costly but transfers efficiently to the robot, while egocentric data can teach broader knowledge about how objects and interactions work. She suggests robots in a known, confined setting could generalize within one to two years, while homes with people and changing environments require stronger planning and safety systems. She describes local execution for small policies and larger models running in the cloud over time.

4. 11:02

   Foundation models and long-horizon tasks

   Huang presents Astribot’s first-generation foundational model, designed to reason about human intent and divide longer tasks into subtasks. In a packing demo, internet-scale data helps the robot connect unfamiliar products to their names. She says future models are aimed at practical tasks such as cooking or preparing a gift, and favors world models that represent motion and task-relevant concepts rather than only pixels.

5. 15:24

   From developers to homes

   Huang describes Astribot’s route from prototypes toward everyday use: first shipping robots to skilled developers, then working with business partners who have valuable tasks for them, and eventually entering homes. She says the next-generation product is part of that effort, while emphasizing that the path is long.