← Back to feed

Boston Dynamics’ Atlas robot carried a full-size mini-fridge — more than 100 pounds — across a workspace using reinforcement learning and proprioception as the primary control signal. Figure AI’s humanoid robot, in a separate demonstration, completed 24 straight hours of autonomous package sorting with zero failures and 30,000 packages processed.

Both milestones landed within 48 hours of each other. They came from opposite ends of the humanoid robotics stack, and together they bracket the two most important unresolved questions in physical AI: can robots handle heavy, dynamic loads, and can they sustain production-grade uptime?

What Atlas Did Differently

Most humanoid load-handling demos are vision-first. The robot perceives the object, plans a grasp, executes. Atlas inverted that priority for heavy objects.

A 100-pound fridge is not a fixed geometry once lifted. It shifts, flexes, and torques against the grip in ways a camera cannot fully anticipate from frame to frame. Atlas used proprioceptive signals — force feedback, balance dynamics, contact pressure — to modulate grip and body posture in real time as the load moved.

Reinforcement learning handled the policy. Domain randomization in simulation — training against varied weight distributions, floor surfaces, and grip geometries — let the policy transfer to hardware without manual tuning. The robot adapted through sensation, not sight.

This matters because heavy-load manipulation is the industrial bottleneck. Factories, warehouses, and construction sites run on tasks that require handling objects humans describe as “awkward” or “heavy” — categories where vision-based robotics has historically failed to close the gap with human workers.

What Figure AI Did Differently

Where Atlas tested manipulation precision, Figure AI tested endurance. A 24-hour continuous run sorting more than 30,000 packages is not a demo. It is an uptime test, and Figure passed it publicly on first attempt.

The 24-hour threshold is operationally significant because most enterprise deployment conversations start there. A robot that cannot sustain autonomous operation for a full work shift cannot replace a shift worker in contract terms. Figure’s robots can now make that commitment.

Figure AI’s approach differs structurally from Boston Dynamics. Where Boston Dynamics develops precise physical intelligence through years of biomechanical research, Figure iterates rapidly on foundation model-driven control and focuses on manufacturing-scale production. The result is a different capability profile: less precision-first, more throughput-first.

The Convergence

Humanoid robotics entered 2026 having already resolved basic locomotion. Honor’s Lightning ran a half-marathon in Beijing faster than the human world record. Unitree shipped a piloted mech. Sony’s Ace robot beat elite table tennis players under official ITTF rules, published in Nature.

The unsolved problems shifted from “can it walk” to “can it work.” Atlas and Figure AI are both answering yes to the work question — from different angles, with different architectures, at close to the same time.

Numbers

• Atlas load: 100+ lbs (mini-fridge, real-world demonstration)
• Atlas method: RL policy trained with domain randomization, proprioception-primary control
• Figure AI run: 24 hours continuous, zero failures
• Figure AI throughput: 30,000+ packages sorted autonomously
• Figure AI target deployment: warehouse and logistics operations

The pace in physical AI has shifted from proving concepts to proving production. Both milestones are production proofs.