Rather than carrying a load, the robot tows it: a single control system manages starting, turning and stopping, whether with human passengers or not, and without the need for any specific adjustments.
A humanoid can move much more than its own weight if it pulls a wheeled vehicle instead of carrying the load. However, it must keep its balance and a stable grip while the cart, the load and the terrain exert changing and unknown forces on it. This is the problem that researchers are tackling with a rickshaw and a Unitree G1 robot.
A teacher who sees everything, a student who guesses
In simulation, a "teacher" model has access to all the privileged information: state of the vehicle, interaction forces, properties of the load. Its behavior is then transmitted to a "student" model which must deduce these dynamics from its history of bodily sensations alone, before being refined by reinforcement learning. Faced with variants with no history or using only history, the policy obtained follows the requested trajectory better while reducing the oscillations of the vehicle, the inclination of the torso and the effort of the engines.
The robot adopts full-body reactions synchronized with its gait to stabilize lateral movements. Pulling redistributes the effort between the joints and is even, in most of the configurations tested, less expensive in relative energy than walking without a load. On real equipment, a single policy links starting, prolonged pulling, turns and stopping with rigid loads or human passengers, up to 115 kg in total.
Read the full article on the original source
Go to the source →