Atlas

Atlas, developed by Boston Dynamics, represents the pinnacle of humanoid robotics research, evolving from its 2013 hydraulic prototype funded by DARPA for disaster response to a fully electric platform unveiled in April 2024. The latest iteration shifts from cumbersome hydraulics to efficient electric actuators, enabling real-world commercial applications, particularly in manufacturing. Standing at approximately 150 cm tall and weighing 89 kg, Atlas leverages a mixture of titanium and aluminum 3D-printed parts for an exceptional strength-to-weight ratio, allowing it to lift payloads up to 18-25 kg and achieve speeds of 2.5 m/s (9 km/h). Its architecture features around 28-50 degrees of freedom (DoF), with full-body configurations offering 50 DoF for locomotion and manipulation, and a Manipulation Test Stand (MTS) variant with 29 DoF focused on upper-body tasks. At the core of Atlas's capabilities is its advanced control system, integrating model predictive control (MPC) for balance and precise motion planning with reinforcement learning (RL) policies. Recent advancements include Large Behavior Models (LBMs), a 450 million parameter Diffusion Transformer architecture developed in collaboration with Toyota Research Institute (TRI). These end-to-end, language-conditioned policies operate at 30 Hz, processing head-mounted monocular camera images, proprioceptive joint encoder data, and natural language prompts to output action chunks for grippers, neck, torso, hands, and feet. Trained via iterative data collection from teleoperation (using VR interfaces and MPC), simulation, and flow-matching objectives, LBMs enable long-horizon tasks like picking, carrying, placing heavy irregular objects (e.g., 22 lb tires), folding robot legs, and navigating workshops while avoiding obstacles and self-collisions. Perception is powered by state-of-the-art vision models: lightweight 2D object detectors for bounding boxes and keypoints, fixture localization fusing 2D keypoints with 3D priors and odometry, and the SuperTracker for 6-DoF pose estimation using render-and-compare on synthetic CAD data. This system handles occlusions, shiny surfaces, and dynamic environments, fused with kinematic and force data for robust state estimation. Hand-eye calibration ensures centimeter-precision manipulation. Deployment history spans research milestones—parkour, backflips, dancing—to practical demos like autonomous engine cover handling. By late 2025, Atlas is in pilot testing with Hyundai for automotive factories, focusing on dull, dirty, dangerous tasks. Integration with Orbit fleet management software supports scaled operations. While still a prototype (not commercially available), Atlas pushes embodied AI frontiers, demonstrating behaviors exceeding human limits in agility and dexterity, with ongoing RL fine-tuning and VLM integration for broader generalization.