SOMA: Stack-of-Objects Manipulation using a Quadruped Robot

Can a legged robot without a manipulator move heavy, unattached stacks of objects to a precise pose with its whole body, without the upper object sliding off or toppling?

Quadrupeds walk well but have no hands. Mounting an arm adds weight and width and limits the load. Pushing with the body suits heavy items, but earlier body-pushing controllers kept the robot at a fixed height, which limits pushes to a flat plane, and optimisation-based planners need accurate models of the object and the floor.

Pedipulation work used one leg as an arm for doors, switches and light objects, usually with the body held still. Arm-on-quadruped systems can grasp, at a cost in footprint and payload. Hierarchical reinforcement learning for whole-body pushing (Jeon et al., 2024) handled single objects at a fixed height and took about 15 hours to train.

SOMA splits control into a 50 Hz walking policy of about 100,000 parameters, which tracks speed, turn rate and a body-height command between 0.1 and 0.3 m, and a 10 Hz manipulation policy that issues those commands. Both use an S5 state-space sequence encoder that remembers recent history, which lets them infer mass and friction from how the box responds. The manipulation policy samples commands from a Beta distribution, which stays within physical limits without clipping. Training runs in three stages (walking, one box, a two-box stack) with box sizes randomised from 0.35 to 0.45 m, mass from 6 to 10 kg and floor friction from 0.3 to 1.5, plus penalties for tilt and for slip between layers.

In simulation, a single 12 kg box, heavier than any box seen in training, was placed successfully about 90% of the time across mid-sized shapes and 82% for tall narrow ones. Training on several sizes beat a single-size baseline by 22.9 to 27.2 points on light 4 kg boxes. Across nine size and mass combinations of two-box stacks, success ran from 71.4% to 90.1% and stack failures from 0.2% to 12.1%, worst for small, light 0.3 m boxes. On the real Go1, with poses supplied by motion capture, the policy ran without fine-tuning; in four trials of the hardest stack setup the position error fell from about 2 m to under 0.05 m in roughly 18 seconds, and the robot dropped its body from about 0.22 m to 0.12 m while pushing.

This is a preprint, and the text does not state the authors' institution. The hardware section lists five setups and plots one, but gives no success-rate table for the real robot. Poses come from an eight-camera OptiTrack system rather than onboard sensors, so the demonstration does not yet work outside a lab. Loads are 4 to 12 kg on flat floors with at most two layers; narrow passages and onboard perception are listed as future work.