The Robot Report Podcast From Teleoperation to Autonomy: Inside Boston Dynamics' Atlas Training
19 snips
Sep 12, 2025 Scott Kuindersma, VP of Robotics Research at Boston Dynamics, leads work on humanoid Atlas and large behavior models. He discusses collecting teleoperation data, training LBMs to generalize bimanual manipulation, using simulation and human demos, and plans to test Atlas in Hyundai facilities. The conversation covers task design, evaluation, and scaling challenges for humanoid learning.
AI Snips
Chapters
Transcript
Episode notes
Workspace-Scale Manipulation Matters
- Boston Dynamics chose the Spot parts task to stress bi-manual, whole-body manipulation across its workspace with realistic environmental constraints.
- The task tested stepping, crouching, and coordinating hands and feet to show generalization beyond tabletop manipulation.
LBMs Mirror LLM Generalization
- Large behavior models (LBMs) analogize LLMs by training on diverse embodied data to generalize many tasks.
- Feeding robot state, camera inputs, and language lets a single model produce closed-loop manipulation policies.
Follow A Four-Stage LBM Pipeline
- Collect teleoperation data, annotate with language, augment visuals, then train and evaluate extensively using simulation and hardware.
- Use simulation rollouts to get statistically significant evaluation before deploying on hardware.