← Paper Digest

WAM-Policy: World Action Models with On-Policy Distillation for Quadrupedal Loco-Manipulation

Junlin Long, Xu Deng, Yiran Wang, Yue Yang, Yang Zhao, Zeyu Zhang

November 2026

The Unitree Go2 presses a button with its front leg, trained only on simulated data.

TL;DR Adapts a world action model pretrained for robot arms into a language-conditioned policy for a Unitree Go2 that walks on three legs and manipulates with a front leg, then distills it to one video and two action sampling steps. On ten simulated tasks the final student succeeds 85.0% of the time (77.0% for the slow transferred policy) while sampling 14 times faster. Deployed on the real robot with no real training data, the distilled Stage 1 student succeeds in 7 of 30 trials.

What they do

An arm-free quadruped can manipulate by standing on three legs and using a front leg as the end effector, but no policy for this follows language. WAM-Policy adapts LingBot-VA, a world action model (WAM) pretrained for dual-arm robots, to a Unitree Go2 that manipulates with its front-left leg. It then distills the model to sample fast enough for closed-loop control and improves it with reinforcement learning.

How it works

  • Cross-morphology transfer. Only the input and output interfaces change. Two body-mounted cameras (ego and top-down) replace the head and wrist views, and a 9-D command for the RAMBO whole-body controller (base velocity, yaw rate, foot target position and force) replaces the 30-D arm command.
  • QLM-Bench. Ten MuJoCo tasks with 1,000 demonstrations. Each task’s command trajectory is drafted by a multimodal model, reviewed by a human and executed in randomized scenes; there is no teleoperation.
  • On-policy self-distillation (OPSD). A LoRA student with 1 video step and 2 action steps regresses onto the 25/50-step teacher’s outputs from the same noise, on contexts from its own rollouts and from cached teacher rollouts.
  • RL and merging. Specialists for pushing a box and opening a door are tuned with a window-level DiffusionNFT, then merged into one few-step student by multi-teacher distillation that routes each context to its task’s teacher.

Results

  • The transferred policy succeeds 77.0% of the time, against 61.7% for a reproduced point-cloud baseline (Leg-Manip), but needs 2.226 s per chunk covering 0.32 s of motion.
  • At 1/2 steps without distillation, success falls to 69.3% (the drawer task to 0%). The OPSD student recovers 76.7% at 0.159 s per chunk.
  • The specialists succeed in 30 of 30 episodes on their tasks. The merged student pushes the box 90% of the time and opens the door 87%, for 85.0% overall.
  • On the real Go2 Air: 7 of 30 trials across three tasks.

Limitations

RL covers only two tasks, each with a hand-designed reward. One robot, one controller and one manipulating leg, with no tests on unseen tasks or instructions. With observation encoding the real control loop takes 0.66 s per chunk, so the robot holds its last command between chunks. Real-robot success is below simulation, and the final student never sweeps the block.

Videos

Simulated and real rollouts side by side, starting with pressing a button.
QLM-Bench tasks in simulation, starting with closing and opening doors.

Figures

From Figure 2: Stage 1. An arm-centric WAM is fine-tuned on quadruped demonstrations synthesized in QLM-Bench. Ego-ce...
From Figure 2: Stage 1. An arm-centric WAM is fine-tuned on quadruped demonstrations synthesized in QLM-Bench. Ego-centric and top-down views replace the head and wrist views, and a 9-dimensional quadruped command replaces the 30-dimensional arm command, which a whole-body controller converts into joint torques. OPSD then distills the high-NFE teacher into a 1/2-NFE student.
From Figure 4: The ten tasks of QLM-Bench, one successful rollout per task, with earlier poses translucent and arrows...
From Figure 4: The ten tasks of QLM-Bench, one successful rollout per task, with earlier poses translucent and arrows marking the intended motion.
From Figure 5: Key frames of the OPSD student pressing a button, kicking a ball and pushing a box. For each task the ...
From Figure 5: Key frames of the OPSD student pressing a button, kicking a ball and pushing a box. For each task the top row is a rollout on the Unitree Go2 Air and the bottom row a rollout in simulation.
Abstract

Arm-free quadrupeds can perform loco-manipulation, walking on three legs while manipulating with a front leg, yet no existing policy for it follows language instructions. World action models (WAMs) have emerged as effective language-conditioned manipulation policies, but are pretrained for robot arms and sample slowly. WAM-Policy turns one WAM into a quadrupedal loco-manipulation policy via two stages. The first, Cross-Morphology Policy Adaptation and Acceleration, adapts the arm-centric WAM to quadrupeds and distills the transferred policy into a student with one video and two action steps by on-policy self-distillation (OPSD). This student, however, is bounded by its teacher, which only imitates demonstrations. The second stage, Task-Specific Improvement and Merging, therefore tunes the teacher by reinforcement learning (RL) and merges the slow specialists into one few-step student by multi-teacher on-policy distillation (M-OPD). On QLM-Bench, ten simulated tasks with 1,000 demonstrations synthesized without teleoperation, the transferred policy reaches 77.0%, and the OPSD student keeps 76.7% while sampling 14 times faster. Task-specific RL and M-OPD lift it to 85.0%. We also deploy the OPSD student on a real Unitree Go2 Air. WAM-Policy thus achieves a stronger quadrupedal policy at a fraction of the inference cost.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.