Tech • AI • Robotics
Boston Dynamics New Atlas Upgrade Makes It Fully Autonomous
8/10TL;DR
Boston Dynamics has equipped the production Atlas humanoid with autonomous battery swapping, a step aimed at making factory robots practical across long shifts while broader advances in simulation-trained control accelerate humanoid deployment in industry, retail and field work.
KEY POINTS
Autonomous battery swaps cut downtime
Atlas can detect low charge, walk to a battery station, remove its depleted pack, install a fresh one and resume work without human help. The swap takes under 3 minutes, compared with about 90 minutes for a normal recharge. For a robot designed for industrial shifts, that sharply reduces idle time and improves utilization.
Designed for factory handling
The electric Atlas stands about 6 feet 2 inches tall, weighs roughly 90 kilograms, and is aimed at lifting, moving and handling materials in changing factory settings. It can sustain loads of about 30 kilograms and briefly lift up to 50 kilograms. Tactile sensing in the hands and a 360-degree camera system support manipulation and situational awareness.
Runtime remains a core economics issue
Under normal operation, Atlas gets roughly 4 hours of runtime, falling to about 2 hours under heavy loads. Without swapping, every recharge could sideline the machine for an hour and a half after only a few hours of work. The new system is therefore less about spectacle than about making continuous industrial operation financially credible.
Simpler hardware supports training and maintenance
Boston Dynamics has simplified the robot’s architecture by using only two actuator types across the body and a symmetrical design for arms and legs. Joints avoid cables running across them, allowing continuous rotation while reducing wear. That simpler layout also makes the machine easier to model accurately in simulation, a key requirement for transferring learned skills to real hardware.
Simulation is becoming the main training ground
The company is using reinforcement learning and large-scale simulation to teach balancing, whole-body coordination and heavy object handling. Modern simulation systems can run the equivalent of millions of training hours in a day, and learned behaviors can reportedly move to the physical robot in about an hour. One test had Atlas move a refrigerator weighing more than 100 pounds after training on substantially lighter loads.
Acrobatic moves serve a practical purpose
Flashy behaviors such as kicks, handstands and backflips are being used to build balance recovery, agility and endurance. Those capabilities matter when a humanoid stumbles while carrying weight or absorbs an unexpected force. The emphasis is shifting from showpiece motion to robust whole-body control for real tasks.
Zest cuts control-system complexity
The RAI Institute and Boston Dynamics developed Zest, short for Zero-shot Embodied Skill Transfer, to train robot motion from motion capture, monocular video and keyframe animation. The policy is trained entirely in simulation and transferred to hardware with zero fine-tuning, using a relatively simple feed-forward network and onboard proprioception. It avoids much of the usual control scaffolding, including contact schedules and long observation histories.
Modest compute, broad results
Each Zest policy took about 10 hours of training, around 7,000 iterations on a single NVIDIA L4 GPU. The framework produced crawling, forward rolls, cartwheels, army crawls and dancing on Atlas, while also handling other body plans including Spot and Unitree G1. Limits remain: the system is confined to flat, non-slippery ground and does not yet generalize to entirely unseen motions.
NVIDIA is pushing a larger-scale model
NVIDIA’s Sonic pursues a similar idea at far larger scale, training one policy on more than 100 million motion frames from roughly 700 hours of motion capture. The largest version uses 42 million parameters and consumed up to 128 GPUs and about 21,000 GPU hours. On Unitree G1, it reached a 99.2 percent success rate across 123 real-world motion sequences and can switch between video, text, music and teleoperation inputs.
Commercial deployment is widening
In China, BYD is deploying the humanoid Xia D in showrooms to greet visitors, explain vehicle models and answer questions, with broader rollout targeted in 1 to 2 years. XPeng aims to scale its Iron humanoid past 1,000 units a month by the end of 2026 and use them as showroom assistants in 2027. The strategy arrives as automakers look for new applications amid softer domestic sales and heavier investment in AI systems.
Policy and field robotics are moving in parallel
On July 28, the US FCC added foreign-produced advanced robotic devices, including humanoids and quadrupeds, to its covered list, restricting new models requiring authorization from being imported, marketed or sold in the United States. At the same time, Deep Robotics has been showcasing the DR02 humanoid on outdoor stairs, uneven grass and around high-voltage infrastructure. The IP66-rated robot carries a swappable 1,440 watt-hour battery and reflects growing investor pressure for proof that humanoids work outside scripted lab demos.
CONCLUSION
Autonomous battery swapping addresses one of the most basic obstacles to humanoid labor: keeping robots productive through a full shift. Combined with faster simulation-to-reality training, it signals a market moving from impressive demonstrations toward deployable machines.
Explain this