Back to articles

WAM-Policy Brings Language-Conditioned Loco-Manipulation to Quadrupeds

3 min read

A quadruped that walks while using one front leg to manipulate an object must coordinate locomotion and contact with its surroundings at the same time. In the setup considered here, the robot has no separate arm: it temporarily relies on three legs for support while its free front leg performs the manipulation. World action models (WAMs) can follow language-conditioned manipulation instructions, but existing models are oriented toward robot arms and can be slow to sample. WAM-Policy explores how to transfer one to a quadruped without retaining the full inference burden.

Adapt first, then improve and compress

The approach has two stages. In “Cross-Morphology Policy Adaptation and Acceleration,” an arm-centric WAM policy is adapted to the quadruped body. The transferred policy then teaches a student through on-policy self-distillation (OPSD), in which the student learns from the teacher’s behavior on the student’s own policy distribution. The paper summary describes a compact student using one video step and two action steps, rather than the teacher’s slower sampling procedure.

The authors identify a ceiling to this first stage: a student distilled from a teacher that only imitates demonstrations is limited by that teacher’s capabilities. Their second stage, “Task-Specific Improvement and Merging,” therefore improves the teacher for individual tasks with reinforcement learning (RL). Multi-teacher on-policy distillation (M-OPD) then combines the slower task specialists into a single few-step student. The design separates two goals: transfer and faster inference first, stronger task performance and consolidation afterward.

Simulation results and what they do not show

Evaluation uses QLM-Bench, a suite of ten simulated tasks, with 1,000 demonstrations synthesized without teleoperation. The transferred policy reaches 77.0% success. The OPSD student records 76.7% while sampling 14 times faster. After task-specific RL and M-OPD, success rises to 85.0%. These results suggest that the initial distillation largely preserves the transferred policy’s benchmark performance while reducing sampling cost, and that task-level reinforcement learning plus policy merging can improve the simulated score.

The authors also deploy the OPSD student on a real Unitree Go2 Air. This indicates that the work extends beyond simulation, but the supplied abstract does not report real-robot success rates, test counts, or a comparison with the simulated results. Deployment alone therefore does not establish how reliably the policy handles real-world tasks. More broadly, WAM-Policy offers a pipeline for adapting language-conditioned manipulation models to a different robot morphology: transfer the policy, distill it for speed, improve task specialists, and consolidate them into a compact student. How well this approach generalizes across tasks and physical conditions remains an important question for further evaluation.

Source: Hugging Face Daily Papers

Comments

Checking sign-in status...

Loading comments...

Related articles