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
cs.RO, cs.CV
2026-11-18
Retargets an arm-pretrained world action model to leg-manipulating quadrupeds: 85.0% success vs 77.0% for the teacher, 14x faster sampling, deployed on a real Unitree Go2.
An arm-free quadruped can walk on three legs and use a front leg as a manipulator: pushing boxes, opening doors, kicking balls. Skipping the onboard arm saves weight and mechanical complexity, but no existing policy for this setting takes language instructions. Prior work learns point-cloud planners, one per task, from demonstrations; language-annotated data for leg manipulation simply does not exist.
Meanwhile, world action models (WAMs), generative policies that jointly predict future observations and actions, have become a standard recipe for language-conditioned manipulation. They are all pretrained for fixed-base arms, and multi-step sampling is slow. The base model here, LingBot-VA, needs 2.226 s to generate one action chunk covering 0.32 s of robot time, about seven times slower than real time. The robot keeps moving while the policy thinks, so actions computed from a stale state no longer fit.
Two stages, each ending with a distillation into a 1/2-NFE student (one video step, two action steps).
Stage 1, cross-morphology transfer and acceleration, touches only two interfaces. On vision, the arm's head and wrist cameras become two body-mounted cameras (a forward wide-angle plus a top-down overhead); both move with the body, so the model must also learn how its own motion changes the views. On action, the 30-D dual-arm command becomes a 9-D quadruped command (planar base velocity, yaw rate, target foot position, target contact force), handed to a frozen whole-body controller (RAMBO) that converts it to joint torques. The VAE, text encoder, and the transformer body initialized from the Wan2.2-5B video generator are kept and fine-tuned on quadruped demonstrations.
Then OPSD (on-policy self-distillation): the teacher is the transferred high-NFE policy itself, frozen; the student trains only a rank-256 LoRA delta with sampling cut to 1/2. Teacher and student see the same context and share one noise draw, the teacher is ahead purely through its 25/50-step sampler, and the student regresses onto its outputs. Training contexts mix the teacher's cached closed-loop rollouts with the student's own streaming rollouts, the latter countering environment-level exposure bias: the student's own errors carry it into states no demonstration covers.
Stage 2 fixes the fact that a distilled student is capped by a teacher that only imitates. For the two hardest tasks, pushing a box (47%) and opening a door (53%), DiffusionNFT trains a LoRA specialist each: from the teacher's own closed-loop states, groups of candidate windows run with one fixed noise pair per window, each window scored by a single reward on the task object's trajectory, candidates above the group mean pulled toward and below pushed away. M-OPD (multi-teacher on-policy distillation) then routes each instruction to its task's specialist and distills the two 2.7 s-per-chunk specialists plus the original teacher into one 1/2-NFE student that continues from the Stage 1 student.
The data is synthesized too: QLM-Bench's 1,000 demonstrations (100 per task, ten tasks) come from an LMM drafting reference trajectories, human review, and retargeting onto randomized scenes. No teleoperation anywhere.
QLM-Bench, ten tasks, Standard100 protocol (100 cases × 3 policy seeds, 300 episodes):
| Method | NFE (video/action) | Per-chunk sampling | Success |
| Leg-Manip (point-cloud baseline, reproduced) | – | – | 61.7% |
| Transferred teacher | 25/50 | 2.226 s | 77.0% |
| Cut to 1/2, no distillation | 1/2 | 0.156 s | 69.3% |
| OPSD student | 1/2 | 0.159 s | 76.7% |
| M-OPD student (final) | 1/2 | 0.159 s | 85.0% |
More numbers worth keeping:
The reusable takeaway outruns any single number: a video-pretrained WAM survives a change of embodiment with only its input and output projections retrained, keeping the trunk weights and language grounding. Cross-embodiment in manipulation usually means a different gripper; here the swap replaces the arm with a load-bearing leg.
OPSD may be the most portable piece: no privileged teacher context, no embodiment assumption, and a clean ablation showing same-noise regression beats distribution matching by a wide margin. QLM-Bench's fully synthesized demonstrations restructure the data cost as LLM drafting plus human review, with no teleoperation.
To be honest about what it is: a systems paper. Transfer, distillation, RL fine-tuning, and multi-teacher merging each have prior work; the contribution is the composition and the arm-free quadruped setting itself.
The authors list four: Stage 2 tunes only two tasks with hand-designed per-task rewards; everything rests on one quadruped, one whole-body controller, one manipulating leg, with no unseen-task or other-robot evaluation; the real-robot closed loop takes 0.66 s per chunk (encoding dominates), twice the 0.32 s the chunk covers, so it is not real time; and the sim-to-real gap is large, 76.7% in simulation against 7/30 on hardware.
A few more from reading closely. The open-door specialist checkpoint comes from round 4, picked on 32 held-out training episodes; during training the reward kept rising while success fell from 19/32 to 6/32 after round 9, so the reward is only a proxy for success there. Specialists bleed into other tasks: the push-box specialist drops drawer-pulling from 29/30 to 21/30. The real robot only sees the Stage 1 student, not the stronger M-OPD one. And the Leg-Manip baseline is the authors' reproduction, with the door task scored 0 for lack of demonstrations (68.6% over the nine tasks with data), which makes the comparison weaker than the headline suggests.