Organizations: University College London · Shanghai Jiao Tong University · The University of Manchester · Chinese Academy of Sciences, Nanjing Institute of Software Technology · Shenzhen University of Advanced Technology · University of Liverpool
World Action Models (WAMs) generate both future video and robot actions, offering two connected outputs for post-training supervision. How can a pretrained WAM learn from a stronger Teacher on the histories it encounters during execution? We present WAM-OPD, which collects Student rollout histories and queries a Teacher for paired video and action targets. The Student learns from both targets while retaining its one-step video and action generation at deployment. Across 12 RoboTwin 2.0 tasks, WAM-OPD improves average success from 33.8% to 65.7%; across four real-robot tasks, it improves average success from 51.4% to 64.6%. With the collected Student histories held fixed, joint video-action supervision achieves the highest observed success on all three ablation tasks, while either modality alone also improves performance. A separate comparison with Teacher-generated histories finds task-dependent differences between the two history sources. These results demonstrate the value of paired video-action supervision for improving WAM policies without increasing their deployed sampling budget.
Figures & tables
Figure 2: WAM-OPD training procedure. The Student interacts with the environment and collects histories hS,i . The Teacher is queried on these histories to produce paired video and action targets (O^iT,A^iT) . The resulting labeled histories form a fixed training set on which the Student recomputes its outputs with the deployed video-to-action computation and optimizes the selected losses.
Policy
Video steps
Action steps
LingBot-VA (Teacher reference)
25
50
Flash-WAM (Student)
1
1
WAM-OPD (Ours)
1
1
Table 1: Video and action sampling steps used by each policy
Figure 3: Representative successful real-robot episodes. Each row shows four chronological frames from one task, from the initial configuration to the completed episode.
Task
Teacher (Reference)
Flash-WAM (Student)
WAM-OPD (Ours)
Δ (PP)
handover_mic
80.0
1.7
35.0
+33.3
put_object_cabinet
51.7
3.3
16.7
+13.3
place_fan
96.7
21.7
50.0
+28.3
place_a2b_left
90.0
11.7
53.3
+41.7
place_a2b_right
90.0
16.7
65.0
+48.3
place_bread_basket
88.3
25.0
60.0
+35.0
Table 2: Task success rates (%) on 12 RoboTwin 2.0 tasks, evaluated over 60 trials per policy and task
Task
Teacher (Reference)
Flash-WAM (Student)
WAM-OPD (Ours)
Δ (PP)
Put Gripper Parts
36.1
5.6
8.3
+2.8
Shake Bottle
94.4
88.9
94.4
+5.6
Insert Screw Driver Tape
66.7
52.8
83.3
+30.6
Put Tape Platform
80.6
58.3
72.2
+13.9
Average
69.4
51.4
64.6
+13.2
Table 3: Real-robot task success rates (%) over 36 evaluation rollouts per policy and task
Figure 4: Effect of supervision composition on task success (%). Each setting uses the same 36-trajectory collection split (24 optimization, 12 calibration) and 60 matched evaluation trials. Only the active loss terms differ. Among these modes, joint supervision yields the highest observed success rate on all three tasks.
Task
Flash-WAM (Student)
Joint / Teacher histories
Joint / Student histories
place_a2b_right
16.7
26.7
65.0
place_bread_basket
25.0
75.0
60.0
place_cans_plasticbox
73.3
100.0
100.0
Table 4: Effect of training-history source on task success (%), evaluated over 60 matched trials per task