From Concept to Production: Humanoid Robotics at Scale

In this talk

1. 6:00

   Collecting useful robot data

   The panel compared teleoperation, human videos, passive video, and data from deployed fleets. Pras said customer data can be difficult to collect and sanitize, so Agility uses targeted robot data alongside pretrained models. Chelsea emphasized collecting large amounts of real robot data across different embodiments, while Deepak described combining human video and simulation for pretraining, then fine-tuning with teleoperation or real-world data.

2. 19:04

   Using simulation and digital twins

   Panelists described simulation as useful for training, testing, and evaluating policies, with different roles across tasks. Deepak said simulation dominates their locomotion pretraining, while real data is more important for manipulation; Chelsea highlighted simulation for testing policies in costly or unusual scenarios. Arnaud described feeding measurements from customer environments back into digital twins to reduce the sim-to-real gap, and shared that reinforcement learning found a way for a wheeled robot to climb stairs by keeping momentum.

3. 31:36

   Choosing a robot brain architecture

   The panel discussed hierarchical and end-to-end approaches to robot control. Chelsea described breaking long tasks into steps, using memory to track progress through occlusions, and accepting higher-level remote guidance. Deepak said Scaled AI trains a hierarchy across multiple embodiments, while Ashok argued that high- and low-level decisions should share information in one model for fast reactions. Russ described co-trained controllers combined with engineered skills to support practical deployment.

4. 46:28

   Making robots reliable over long deployments

   For robots operating continuously, panelists pointed to reliability, safety, and fleet coordination as central needs. Agility already connects Digit to warehouse and manufacturing systems, chargers, conveyors, and mobile robots through a cloud orchestration platform. Panelists also described long-running agents that can retry tasks and adapt, and raised the possibility of sharing successful recoveries across robot fleets. The acceptable failure rate depends on the application.