Digit Humanoid Nails a 65-Pound Deadlift and Reveals How Agility Trains Its Robots

Digit is seen performing deadlifts with a 65-pound weight in the center of a lab. Agility Robotics shared the video a few days ago, and to be honest, the robot maintains a fairly steady balance and completes the task from beginning to end. Someone mentions that the new version can lift significantly more weight than the previous one, while another laughs about how it can run all day without stopping.

The engineers designed the test so that Digit had to work harder than usual. Every additional pound it must lift causes the robot to modify its entire body at simultaneously, including its arms, legs, torso, and everything else. The system must keep the weight centered and avoid tipping over, therefore the legs, arms, and rest of the robot must all function together. These actuators and joints can withstand repeated load without breaking down. Digit’s video simply shows the robot grasping the weight, rising up, then effortlessly placing it down repeatedly in a standard indoor location built for people.

Simulation is where all of the training takes place, because before it touches a real weight, an engineer creates a digital copy of the same thing in a virtual world. Then they anticipate what will happen when the weight shifts. The grip pressure remains constant, with no slipping or lowering. Any changes to the robot’s equilibrium are registered extremely instantly. The policy learns the perfect lift in the simulated environment with no complications before being transmitted directly to the real robot. When you see the real robot perform it, it looks fairly natural because it has already handled every potential variable thousands of times in the simulation.

Engineers chose deadlifts for the test because the movement requires complete body control. A simple arm raise would not put the hardware under the same level of stress. By incorporating weight into the simulation loop, the team is able to handle balancing changes that a pre-programmed script cannot handle alone. As a result, Digit lifts consistently, with no wobbling or resets. This method is easily adaptable to other objects or larger loads in future tests.

Digit was built by Agility to manage long, repetitive jobs that wear people out, such as working in factories or warehouses where you must squeeze into tight spaces, pick up oddly shaped goods, and continue without taking a break. This deadlift test demonstrates Digit’s ability to lift weight on ordinary floors while remaining steady, which is ideal for picking up boxes, carrying tools, and stacking things in human-designed places.

It also illustrates how far they’ve come in teaching robots to perform physical tasks. Whole-body synchronization was originally a nightmare, with hand-tuned code for each joint angle. But now they can simply train a policy in simulation that adapts on the go. Digit detects weight using its sensors, corrects itself in real time, and completes the lift without assistance, while the hardware can keep up because the training has already taught the actuators and joints to be more durable.
[Source]