Evaluation & Deployment

Only the intent estimator and the student actor are deployed, next to the low-level locomotion controller. They need the arm joint states and nothing else: no force-torque sensor, payload pose or partner tracking.

Play in simulation

In legged_gym's play.py, run the intent estimator on the observation history at every step, append its output to the observation, and run the student actor on the result.

Metrics

The paper evaluates all methods with fixed payload masses of 2, 4 and 8 kg under the same wrench profiles:

• Tracking error: mean ‖c_xy − v_xy‖ (m/s) and |c_z − ω_z| (rad/s), using the wrench as the velocity reference.
• Estimation error: mean absolute error between ŵ and c, scaled by 40 N and 10 N·m.
• End-effector force and torque: magnitude of the wrench the end-effector exerts on the payload during steady-state transport. Lower means the robot yields to the partner instead of resisting.
• Intent alignment: cosine similarity between the applied force and the end-effector velocity, F^⊤v_ee / (‖F‖ ‖v_ee‖).

Export

Export the intent estimator and the student actor together as one TorchScript or ONNX module that maps the observation history to the action. Also returning the estimated wrench is useful for logging: multiplied by (40, 40, 10), it is the partner wrench in N and N·m.

Multi-robot teams

The single-agent student policy transfers to teams without retraining. Each robot runs its own copy on its own proprioception, with no communication or central planner. The rigid shared payload combines the leader's and the teammates' wrenches, and each robot perceives the result through its arm joints in the same wrench space it was trained on.

• In simulation, spawn several robots and one shared payload per environment, with each end-effector holding the payload. We evaluated teams of 2 to 4 robots with 12-28 kg payloads of different shapes (sofa, desk, barrel).
• In the real world, a human leader and two robots carried an 8 kg long box.

Other embodiments

The same formulation works for a quadruped without an arm. The high-level policy then observes a history of leg joint states instead of arm joint states and outputs only the 3-D base command.

Limitations

• The high-level policy has no exteroception. It follows the leader and does not avoid obstacles, so walls or narrow passages can lead to collisions and falls if the leader does not steer away.
• Intent is a planar wrench (F_x, F_y, Θ_z). Lifting, tilting or full 6-D interaction is not modeled.