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.
Figures & tables
Figure 2: Stage 1: Cross-Morphology Policy Adaptation and Acceleration. Stage 1.1 ( Sec. 3.1.1 ) fine-tunes an arm-centric WAM on quadruped demonstrations synthesized in QLM-Bench ( Sec. 4 ). Visual Adaptation replaces the head and wrist views (images from Mobile ALOHA [ 13 ] ) by the ego-centric and top-down views, Action Adaptation replaces the 30-dimensional arm command by the 9-dimensional quadruped command, and the whole-body controller C converts each command into joint torques. Stage 1.2, OPSD ( Sec. 3.1.2 ), distills the transferred policy, as the high-NFE teacher πT , into the 1/2-NFE student πS ; the generation time counts the sampling of video and action only. Bottom right shows the closed loop of the policy on the task “Press the button.” Imagined views and rendered demonstrations are illustrations.
Figure 3: On-Policy Self-Distillation (Stage 1.2, top) and Stage 2 (bottom). Top left (Training Context): training contexts mix the teacher cache, which the frozen teacher collects in closed loop before training, with the contexts that the student reaches in its own streaming rollouts ( H∼dˉ ). Top right (Target & Loss): on each context, teacher and student sample from the same noise, and the student’s one-step video stream and two-step action stream are regressed to the targets of the teacher’s high-NFE sampler, with a stop-gradient on the plan that conditions the student’s action. Bottom: Stage 2.1 tunes the OPSD teacher by reinforcement learning into a specialist teacher for each selected task ( Sec. 3.2.1 ), and Stage 2.2, M-OPD, distills the teachers, routed by task, into one 1/2-NFE student that starts from the OPSD student ( Sec. 3.2.2 ). Plan images and rendered actions are illustrations, not model outputs.
Figure 4: The ten tasks of QLM-Bench. One successful rollout per task, with earlier poses translucent; arrows mark the intended motion.
Method
NFE
Time (s)
Fridge
Oven
Valve
Drawer
Door
Press
Basket
Kick
Push
Sweep
All
Seeds
Leg-Manip [ 17 ]
–
–
93
90
93
70
0
80
80
87
17
7
61.7
–
Stage 1: OPSD
Teacher
25/50
2.226
100
100
100
97
53
77
100
97
47
0
77.0 ± 4.0
77/73/81
Student w/o distillation
1/2
0.156
100
100
47
0
93
93
100
97
60
3
69.3 ± 1.2
68/70/70
Student (Ours)
1/2
0.159
100
100
100
97
37
83
100
100
50
0
76.7 ± 4.7
82/75/73
Stage 2: M-OPD
Table 1: Success rate (%) on QLM-Bench Standard100. WAM-Policy rows use 300 episodes (100 cases × 3 policy seeds). ± is the standard deviation of the ten-task mean over policy seeds 42, 43, and 44, and Seeds gives the successes out of 100 episodes for each. NFE gives the video/action sampling steps. Time is the per-chunk sampling time of the two streams on an RTX 5070 Ti, without encoding the observations and with the LoRA update of each student merged ( Sec. G ); a chunk covers 0.32 s of robot time. Leg-Manip is averaged over three training seeds, each evaluated on the 100 cases with evaluation seed 42, so it has no results per seed. Each specialist teacher is evaluated on all ten tasks but used by M-OPD only on its own task ( Eq. 4 , Sec. F ). Shaded rows are our method.
Table 5
Figure 5: Key frames of real-robot and simulated rollouts of the OPSD student on the three tasks of Tab. 7 . For each task, the top row (Real) is a rollout on the Unitree Go2 Air and the bottom row (Sim) a rollout in QLM-Bench; time runs from left to right.
Figure S1 : Specialist teachers during Stage 2.1. Solid line (left axis): reward gain over the OPSD teacher on held-out start states after each round, paired by start state and noise, with a band of ±1 standard error over scenes; the two rewards measure different quantities (the box reward of Eq. S9 in mm, the offset of the foot on the door handle in cm). Squares (right axis): closed-loop success over 32 episodes after rounds 4, 9, and 14; the dotted line is the success of the OPSD teacher. For opening the door, the success falls from 19 of 32 episodes after round 4 to 6 after round 9 while the reward gain stays positive, so the reward is only a proxy of success there; the specialist teacher of Tab. 1 is the update after round 4 ( Sec. F ).
Variant
NFE
Updates
Fridge
Oven
Valve
Drawer
Door
Press
Basket
Kick
Push
Sweep
All
Seeds
Training contexts ( Tab. 7 )
Teacher cache only ( η=1 )
1/2
300
100
100
100
100
33
90
100
100
47
0
77.0 ± 3.5
79/73/79
Streaming only ( η=0 )
1/2
300
100
100
100
97
50
93
100
100
23
0
76.3 ± 4.0
80/72/77
OPSD (Ours, η=0.5 )
1/2
300
100
100
100
97
37
83
100
100
50
0
76.7 ± 4.7
82/75/73
Longer training, policy seed 42 ( Tab. 7 )
Teacher cache only ( η=1 )
1/2
300
100
100
100
100
60
90
100
100
40
0
79.0
79
Table S1 : Ablations on QLM-Bench Standard100 per task (success rate, %), with all rows of the ablation tables of the main paper. The protocol, ± , and Seeds are as in Tab. 1 , and Updates counts the updates of the distillation. Groups marked with policy seed 42 use that seed only (100 episodes, 10 per task), so they have no ± and one entry of Seeds.
Figure S2 : Distillation loss. The held-out loss is evaluated every 50 updates. Top: held-out loss of the students of the three training-context sources of Tab. 7 , on the same 64 held-out contexts, with a square-root horizontal axis; the dotted line marks 300 updates, after which the students are evaluated. The held-out contexts are drawn from the teacher cache, so the curves do not rank the sources: they disadvantage Streaming only, which trains on no cached contexts. Bottom: the M-OPD student ( Sec. 3.2.2 ), which starts from the OPSD student and learns from the specialist teachers routed by task. The light line is the training loss of every update, computed on the contexts sampled for that update rather than on the held-out contexts; the dark line is its centered moving average over 25 updates, with shorter windows at both ends; the circles are the held-out loss. The held-out contexts of M-OPD come from its own teacher cache, so its loss is not on the scale of the top panels.
LIBERO-Pro2.0
Policy
BASE
Static
Runtime
Long
Teacher (high-NFE) †
–
63.7
79.6
98.5
Student w/o distillation
95.0
64.4
68.5
91.8
Teacher cache only ( η=1 )
100.0
69.2
70.8
96.4
Streaming only ( η=0 )
93.3
70.2
72.4
–
OPSD (Ours)
98.3
68.0
72.3
98.2
Table S2 : Success rate (%) of an arm WAM on LIBERO. LIBERO-Pro2.0 reports the official group rates; BASE has 60 episodes. Every student is evaluated after 300 updates, and Streaming only is not evaluated on LIBERO-Long (–). † Published numbers under different protocols, not run by us: LIBERO-Long from Li et al . [ 24 ] , and static and runtime values are our means over the perturbations in the table of Zhou et al . [ 73 ] , which lists no BASE rate for these tasks.
Variant
Updates
BASE
Static
Runtime
Teacher cache only ( η=1 )
300
100.0
69.2
70.8
Streaming only ( η=0 )
300
93.3
70.2
72.4
OPSD (Ours)
300
98.3
68.0
72.3
Collected once
300
96.7
70.2
72.3
SFT only (no teacher)
692
90.0
65.9
69.7
Table S3 : Ablation studies on LIBERO-Pro2.0: success rate (%). SFT only receives 692 updates, whose supervised chunks match 2,000 updates of the teacher-cache-only variant, 6.7 times the budget of the distilled students.
LIBERO-Pro2.0
Policy
Updates
BASE
Static
Runtime
Long
Teacher cache only ( η=1 )
300
100.0
69.2
70.8
96.4
Teacher cache only ( η=1 )
2,000
100.0
68.9
73.5
96.8
OPSD (Ours)
300
98.3
68.0
72.3
98.2
OPSD (Ours)
900
95.0
69.5
72.1
97.2
Table S4 : Success rate (%) on LIBERO after longer training. The 300-update rows are those of Tab. S2 . Table 7 gives the corresponding QLM-Bench results.
Figure S3 : Real-robot setup of Sec. 5.5 . Left: the Unitree Go2 Air, which carries the two cameras and an RDK X5 board and manipulates with its front-left (FL) leg, connects to the workstation by Ethernet. Right: WAM-Policy takes the ego-centric and top-down views and the task prompt; the video stream predicts the plan, the latent of the next observation, and the action stream infers the action chunk from it by inverse dynamics. The whole-body controller RAMBO [ 9 ] executes the chunk on the robot with an RL policy and joint PD control.
Figure S4 : Fine-tuning losses of the transferred policy (Stage 1.1), on a logarithmic scale. Light: training loss of each update; dark: its centered moving average over 100 updates, i.e . one pass over the 800 training demonstrations; black dashed: validation loss on 10 held-out demonstrations per task, every 100 updates.
Comparison
Δ
+/−
p
300 episodes
tc vs. W/o
+7.7
54/31
0.017
str vs. W/o
+7.0
49/28
0.022
OPSD vs. W/o
+7.3
53/31
0.021
tc vs. T
0.0
12/12
1.0
str vs. T
−0.7
11/13
0.84
Table S5 : Paired comparisons on QLM-Bench. Δ is the success rate of the first policy minus that of the second, in points; +/− counts the pairs in which only the first or only the second policy succeeds; p is the exact two-sided McNemar test on these counts. T: Teacher, the OPSD teacher; W/o: Student w/o distillation; OPSD: our student, the Student (Ours) of Tab. 1 ; M-OPD: the M-OPD student of Tab. 1 ; tc, str: Teacher cache only ( η=1 ) and Streaming only ( η=0 ). Students are evaluated after 300 updates unless a superscript gives the number of updates. 228 of the 300 episodes of the M-OPD student run under the second evaluation configuration of Sec. G ; its 60 episodes of pushing the box and opening the door run under the configuration of the other rows.
Comparison
Δ
+/−
95% CI
p
LIBERO-Pro2.0, 2,520 episodes
tc vs. W/o
+3.8
237/143
[1.5, 6.0]
2×10−6
str vs. W/o
+4.9
249/127
[2.9, 6.8]
3×10−10
OPSD vs. W/o
+3.9
232/136
[1.7, 6.0]
6×10−7
OPSD vs. tc
+0.1
127/125
[ −1.1 , 1.3]
0.95
str vs. OPSD
+1.0
142/116
[ −0.3 , 2.5]
0.12
Table S6 : Paired comparisons on LIBERO. Δ , +/− , and p are as in Tab. S5 , and 95% CI is a bootstrap interval of Δ . W/o: Student w/o distillation; OPSD: the OPSD (Ours) of Tab. S2 ; tc, str, once: Teacher cache only ( η=1 ), Streaming only ( η=0 ), and Collected once; SFT: SFT only (no teacher). Students are evaluated after 300 updates unless a superscript gives the number of updates.