Brett Adcock Interview | Helix 2.5, the Holy Grail of Robotics and Figure 4

In this talk

1. 0:26

   What general-purpose home robotics requires

   Adcock describes a humanoid that can respond to spoken requests and complete household jobs independently and safely. He says it needs whole-body autonomy built with deep learning, plus the ability to work in unfamiliar homes with different layouts, appliances, and objects.

2. 2:21

   Four stages of humanoid development

   Adcock lays out four stages: capable hardware, an end-to-end autonomy system, broader robot intelligence, and manufacturing at scale. He says the first two depend on sustained engineering, while intelligence and production can grow with investment in data, compute, and manufacturing capacity.

3. 6:45

   Scaling data and evaluating Helix 2.5

   Adcock says Figure tested Helix 2.5 in rental homes and observed better performance as training data increased. The evaluation focused on three complete tasks: picking up living-room toys, making a bed, and folding towels. He describes Index as a worldwide human-data collection effort, reporting 107,000 active users that week and about 50 minutes of incoming video per second.

4. 10:05

   Compute plans and the limits of current results

   Adcock says new Index videos are prepared for training and describes a plan with Nscale involving up to 100,000 NVIDIA Vera Rubin GPUs. He says the home evaluations suggest a promising scaling trend, but stresses that long-horizon work in arbitrary homes is not yet solved and that the reported task set was deliberately narrow.

5. 29:46

   Figure 4 and manufacturing progress

   Adcock declines to share design details about Figure 4, but says it will be a major hardware step beyond Figure 3. He reports that Figure had built its 10,000th Figure 3 the previous month and describes how on-site manufacturing lets engineers and production teams iterate together.