Agility’s Digit Humanoid Robot Learns to Dance

Agility Robotics has spent years developing their robot Digit with the purpose of enabling it to perform real-world tasks in warehouses where people move heavy boxes and totes all day. The company just released a short video showing Digit performing some slick dance routines, with excellent coordination and steady balance throughout.

Digit stands 5 feet 9 inches tall and weighs approximately 140 pounds. Its legs are modeled like an ostrich’s, and they bend backwards at the knee, similar to Cassie, Agility’s earlier robot. This construction provides Digit with extremely solid footing and allows it to hunch down to reach shelves or navigate narrow passageways that are primarily built for humans. Each arm contains several joints that allow it to lift and carry up to 35 pounds at a time. Inside the torso, a number of cameras and a lidar sensor are constantly examining the environment. They have two major computers handling the planning side of things, while a separate unit is processing the vision data. The battery lasts around 4 hours between charges, and when it runs low, the robot can simply roll itself to a docking station.

Agility Robotics was founded in 2015, when it branched off from Oregon State University. Co-founders Jonathan Hurst and Damion Shelton had been developing dynamic walking robots in academic labs and were eager to see the technology applied in the real world. Their first platform, Cassie, demonstrated that a two-legged robot could run, jump, and even recover from being pushed without colliding with its face. Digit debuted in 2019, and it was far more practical, with a full upper torso, arms, and all the sensors required to do real labor. The company currently makes the robots in Salem, Oregon, because they employ largely American-made parts and want to maintain their supply chains looking reliable. There are currently some quite large warehouses running fleets of Digit units that transport thousands of plastic totes per shift.

The dance routines you see are the result of a training approach known as sim-to-real reinforcement learning. What they do is collect several types of raw motion data, some from motion capture suits, some from animated sequences, and some from plain old teleoperation. They then feed all of that data into extremely complex computer simulations, where virtual Digits can practice the dancing steps thousands of times. The software has a few tricks in its sleeve, as it rewards Digits for maintaining their balance and making smooth transitions, while slapping them on the wrists if they fall over or become jerks. All of this happens overnight in the simulation, and once completed, the control policy is transferred to the real-world robot. The real-world Digit then goes through a few test runs to fine-tune its movements, and the entire process is now so quick that new skills and tricks develop seemingly out of nowhere.