Spring 2024 GRASP SFI Eric Jang, 1X Technologies, “Data Engines for Humanoid Robots”

In this talk

1. 3:42

   Why humanoids and physical AI matter

   Jang argues that software can coordinate information but cannot perform physical work such as packing boxes or cleaning homes. He says human-shaped robots suit homes built around human movement, while wheeled bases can struggle with clutter and reach. He also points to compliant, quasi-direct-drive hardware as a way to make contact safer and learning more forgiving.

2. 18:48

   Safety, capability, and cost

   Jang says home robots need broad capabilities, safe interactions, and affordable, repairable hardware. He contrasts this with deploying systems on tasks where failures carry especially high stakes, and says lower-stakes home use may allow more iteration. He expects humanoid costs to fall, citing examples ranging from a $90,000 Unitree H1 to a roughly $20,000 target mentioned for Optimus.

3. 22:11

   A data engine built around teleoperators

   1X deploys autonomy, identifies failures, collects data from those situations, retrains, and repeats. Teleoperators both gather data and train models, allowing them to choose useful examples rather than follow fixed collection instructions. Jang says this can improve sample efficiency and data quality, and notes that uploaded fleet data is only a fraction of what robots have collected.

4. 25:07

   Learning with generative models

   Jang describes neural networks taking on more of the traditional perception, planning, and control pipeline. Examples include a diffusion model that imagines a one-second future image for door opening and diffusion policies that generate actions from observations and goals. 1X also trains learned gaits for Neo in simulation, while Eve currently uses a whole-body controller for balance.

5. 30:29

   Open challenges and evaluation

   Jang sees general-purpose robotics as promising but says policies still need to generalize across locations, objects, and equipment. He identifies evaluation as a major challenge and suggests continuous measures, such as repeated door openings before failure, alongside broader task sets. He also discusses distribution shift, adapting to home environments, and balancing new data against the cost of retaining and combining older datasets.