Hardware Podcasts

From Teleoperation to Autonomy: Inside Boston Dynamics' Atlas Training

The Robot Report Podcast · 12 Sept 2025 · 1h 05m

AI in practice: 4 real uses of AI described

What this episode covers

Scott Kuindersma, VP of Robotics Research at Boston Dynamics, explains how the company trains the Atlas humanoid robot using Large Behavior Models (LBMs) and teleoperation data. The team collected 20 hours of teleoperation data to enable Atlas to perform complex bi-manual manipulation tasks, such as picking and placing Spot parts, while adapting to real-time disturbances. Kuindersma details the four-stage pipeline of data collection, annotation, model training, and evaluation, highlighting the critical role of simulation data and human demonstrations in generalizing these skills. The episode concludes with plans to test Atlas in Hyundai facilities and further explore AI-centric dynamic behaviors.

AI in practice

Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.

Engineering & simulationLarge Behavior Models (LBMs)20 hours of teleoperation data

Boston Dynamics used teleoperation data and machine learning to train models that allow the Atlas robot to generalize manipulation tasks and react to environmental disturbances in real time.

“we're using teleoperation as a way to demonstrate examples of the behavior that we want the robot to perform, and then we're using, basically machine learning to take those that data and turn it into a model that can perform many of the tasks that we demonstrated.”

Scott Kuindersma · Hear it at 31:17

Engineering & simulationLarge Behavior Models (LBMs)20 hours of data; cloud computing cluster

The team collected 20 hours of human teleoperation data, annotated it with language, and trained a neural network on a cloud computing cluster to create a policy that controls the robot's closed-loop manipulation.

“So for the spot workshop task, I think we, we probably had on the order of, uh, 20 hours of data. … we take that data. it's annotated now. We run training. This is typically done on a, you know, a cloud, uh, computing cluster, lots of GPUs.”

Scott Kuindersma · Hear it at 37:03

Engineering & simulationSimulation

The team leveraged simulation to perform many evaluation rollouts, allowing them to generate statistically significant statements about policy performance differences without needing to run every test on physical hardware.

“It's also really useful to leverage simulation as a way to get many, many, um, evaluation rollouts. So our collaborators TRI had a paper that they released a month ago, I think did a great job of this and just leveraging lots and lots of simulation rollouts so they can get statistically significant statements about the difference in performance…”

Scott Kuindersma · Hear it at 37:03

Engineering & simulationSimulationimproved performance on physical robot

Incorporating simulation data as part of the training mixture with real robot data was found to improve the performance of the policy on the physical robot.

“We've actually found, in our experience, that, um, and other groups have too, that incorporating simulation data as part of the training mixture actually improves performance on the physical robot, which is really exciting.”

Scott Kuindersma · Hear it at 52:50

Most relevant to

early-stage companies

Topics

Mentioned

Companies
Boston Dynamics, TRI, Toyota Research Institute, Serve Robotics, Phantom Auto, Voices AB, Waymo, AWS, iRobot, Hyundai, Universal Robots
Products
Atlas, Spot, Robel Maker
Also
ISO 10218

Related episodes from other shows

Other shows on the same ground, closest first.

  • Robot Talk · 24m

    Episode 146: Embodied AI on the ISS - Jamie Palmer

    Jamie Palmer, co-founder and CTO of Icarus Robotics, explains how his company is building free-flying dexterous mobile robots to perform routine and risky tasks in low Earth orbit, starting with the ISS. Palmer details the engineering challenges of controlling a 20-degree-of-freedom robotic arm in microgravity, where moving an arm creates equal reaction forces that destabilize the base, requiring advanced flight control similar to underwater robotics. He highlights the role of recent AI breakthroughs, such as Google's RT-2 and Open X Embodiment, in enabling general-purpose manipulation models trained on teleoperation data, and outlines the company's path from remote testing to a 2027 mission with Voyager Space.

    Robotics & autonomyEmbedded firmware
    7 chapters

  • Robot Talk · 20m

    Episode Thirty-Nine: Maria Bauza Villalonga

    Maria Bauza Villalonga, a research scientist at DeepMind, discusses her work on robotic generalization and the path toward artificial general intelligence (AGI). She explains how probabilistic models allow robots to simulate outcomes and transfer skills across tasks, drawing on her PhD research at MIT’s Manipulation Lab and the Amazon Robotics Challenge. The conversation covers the importance of tactile sensing for force control, the role of embodiment in achieving true intelligence, and the scaling advantages DeepMind provides for robotics research.

    AI & machine learningMechanical engineering
    7 chapters

  • The Next Byte Podcast · 23m

    177. Did A Shrimp Fry This Rice? No, A Robot Did

    Daniel and Forbode discuss Stanford University's 'Mobile Aloha' platform, a bimanual robot designed to learn household tasks through imitation learning rather than extensive simulation. They highlight the platform's ability to generalize from a static dataset of 825 operations and improve success rates by 90% after roughly 50 human-led demonstrations. The hosts emphasize the significance of reducing the cost of such robotic systems from hundreds of thousands to around $30,000, making general-purpose robotics more accessible for future commercial and domestic use.

    AI & machine learningRobotics & autonomy
    7 chapters

  • The Next Byte Podcast · 15m

    229. Teaching Robots to Play Badminton

    The hosts discuss ETH Zurich's development of a quadruped robot playing badminton, highlighting the integration of Anymal robots and DINA robotic arms. They explain how this system uses a single reinforcement learning control policy to coordinate leg movement and arm swinging in real-time, stress-testing perception and planning capabilities. The episode connects this research to practical applications in surgical robotics, search and rescue, and collaborative manufacturing (cobots), emphasizing the transition from academic theory to industrial practice.

    Robotics & autonomyEmbedded firmware
    7 chapters

  • Robot Talk · 29m

    Episode 74: Teaching robots to grasp - Edward Johns

    Edward Johns, Director of the Robot Learning Lab at Imperial College London, discusses the challenges of robotic manipulation, arguing that the bottleneck is currently AI and control rather than hardware. He explains his lab's use of imitation learning, where robots learn by being physically guided, and details the Dali Bot project, which uses AI image generators to help robots visualize task goals. Johns outlines applications ranging from warehouse sorting to flexible manufacturing and long-term domestic robots, while noting the significant data scarcity in robotics compared to other AI fields.

    Robotics & autonomyAI & machine learning
    7 chapters

The summary, chapters and topics are generated from the episode’s audio by this directory, not written by The Robot Report Podcast. The episode itself belongs to its publisher.