Tech • AI • Robotics • Game
Boston Dynamics New Atlas Upgrade Just Breaks the Limits of Movement
7/10TL;DR
Boston Dynamics turned Atlas into a wheeled research platform with a one-part retrofit to stress-test its electric motors, validate simulation-trained control, and feed the results back into humanoid robots for factory use.
KEY POINTS
A minimal hardware change
Atlas was converted from a walking humanoid to a rolling one by replacing its feet with off-the-shelf scooter rims mounted directly to the existing ankle motors through a single custom adapter. No other hardware changes were required. The swap was possible because Atlas was designed with modular parts for high-volume production and servicing, including deployments planned at Hyundai facilities.
Why wheels were added
The wheeled version is not a product and is not intended for sale. Its purpose is to push Atlas’s electric motors harder than walking does. While walking loads leg motors in short bursts, rolling forces them to deliver high torque at speed during launches, turns and stops while supporting the robot’s body weight and potential payloads.
Simulation drove the learning
In software, each wheel was treated as an ankle joint allowed to spin continuously, avoiding a full rebuild of the robot model. Atlas learned its rolling behaviors through reinforcement learning, repeatedly practicing in simulation and being rewarded for successful motion. The central challenge was ensuring those simulated skills would transfer cleanly to the physical robot.
Near-perfect transfer to the real robot
Engineers first measured the behavior of the real motors and tuned the simulated motors to match them closely. They also kept low-level motor settings consistent across software environments. That approach produced a highly reliable sim-to-real transfer, with the wheeled behaviors working on the physical Atlas with little apparent mismatch.
Existing control systems were reused
Rather than writing a new controller from scratch, the team kept Atlas’s whole-body control framework and changed only three elements: the rewards used during training, the command ranges the robot practiced, and the way it managed posture. The result was stable, steerable rolling without hand-scripted motions. Boston Dynamics combines classical physics-based control with learned control, using the latter to adapt to bumps, uneven footing and other conditions that are difficult to model in advance.
New movements emerged from learning
Some behaviors were based on motion-capture data from humans, especially clips with heavy foot sliding. In a sideways dash, Atlas did not imitate the human’s dragging and hopping steps but instead learned to steer and roll laterally. In a 360-degree spin jump, it landed the move on its first real-world attempt and added its own pre-takeoff lower-body windup, a motion unavailable to a human body without wheels.
A broader industry debate
The experiment highlights a central question in humanoid robotics: wheels or legs. Wheels are faster and more efficient on smooth industrial floors, while legs better handle stairs, clutter and spaces built for people. Hexagon designed its Eon humanoid around wheels, and BMW is bringing it into its Leipzig plant, while companies such as Agility Robotics continue to emphasize legged machines like Digit.
Lessons are feeding back into standard Atlas
The motor models and whole-body learning methods developed in the wheeled trials are being applied to the regular legged Atlas. Boston Dynamics is preparing broader Atlas testing at Hyundai sites and continuing development of a new four-finger hand. Another priority is fault response: improving how the robot reacts when a motor fails, a key safety issue for machines working near people.
Parallel efforts are rethinking robot bodies
Separate research is exploring softer, more human-like mechanics. A University of Tokyo team led by Kento Kawaharazuka built life-size robotic legs using TPU to 3D-print muscles, tendons and ligaments by changing internal lattice patterns. The legs managed alternating steps with external suspension, but durability and unsupported walking remain open problems, even though worn parts could be cheaply reprinted.
Fabric muscles for wearable robots
At the Korea Institute of Machinery and Materials, a team led by Dr. Chul Hun Park developed “fabric muscle” from shape-memory alloy yarn woven into cloth. About 0.3 ounces of the material can lift roughly 33 pounds, and a wearable system assisting the shoulder, elbow and waist weighs under 4.5 pounds. In tests it reduced muscle effort by more than 40%, while a shoulder version weighing under 2 pounds improved movement in muscular dystrophy patients by more than 57% at Seoul National University Hospital.
CONCLUSION
The wheeled Atlas shows that the most important advances in humanoid robotics may come less from dramatic hardware overhauls than from better motors, more accurate simulation and controllers that transfer reliably to the real world. At the same time, parallel work on printed tissues and fabric actuators suggests the future of robots may split between powerful industrial machines and softer wearable systems built to move with the human body.
Ask a question