ODEAction: From Predicting Chunks to Evolving Actions
Authors: Qishen Wang, Pengxiang Ding, Shifeng Bao, Mingyang Sun, Hengtao Li, Yi Gao, Baohua Yin, Yang Liu, Jiajun Fu, Jiacheng Liu, Changjing He, Peiyang Li, Runze Suo, Yan Wang, Xiaobo Wang, Yonghao Dang, Cheng Chi
Organizations: Symbiosis Robotics · University of Science and Technology of China · Zhejiang University · Westlake University · Renmin University of China · Tsinghua University · Beijing University of Posts and Telecommunications · Fudan University · Shenzhen University of Advanced Technology · Central University of Finance and Economics
Vision-language-action (VLA) models typically generate actions in chunks, dividing continuous robot execution into separately predicted action chunks, which can introduce discontinuities at chunk boundaries. We propose ODEAction, a chunk-free action generation paradigm that replaces action chunk prediction with Continuous Action Evolution. We use an ordinary differential equation (ODE) to describe the continuous evolution of actions, and define the vector field of this ODE as the Action Transition Field to characterize how actions evolve with respect to evolution time. During inference, the model takes the current action as the initial condition and numerically solves the ODE defined by the learned Action Transition Field, thereby recursively generating subsequent actions with explicit dependencies. ODEAction is implemented as a modular action head that can be integrated with different vision-language backbones. Experiments on LIBERO, EBench, RoboTwin, and real-robot tasks show that, compared with action-chunk-based baselines, ODEAction improves both task performance and trajectory quality, with particularly pronounced gains on fine manipulation tasks.
Figures & tables
Figure 1: From action-chunk prediction to continuous action evolution. (a) Base Policy uses action-chunk prediction and is prone to abrupt action changes at chunk transitions, whereas ODEAction’s continuous evolution ensures that action trajectories are generated continuously. Comparisons of real-robot task execution and the corresponding end-effector trajectories illustrate the differences in execution smoothness and trajectory continuity between the two methods. (b) Representative results on EBench simulation and real-robot tasks: compared with Base Policy, ODEAction increases the task success rate (SR) and reduces the end-effector mean absolute jerk (EE-MAJ).
Figure 2: Learning the action transition field and recursively generating actions. (a) Discrete demonstrated actions are interpolated into a continuous trajectory. Target transition vectors are constructed by combining progression along the demonstration with deviation correction, inducing a target action transition field F∗ in the trajectory neighborhood. (b) During training, the action transition field ( Fθ ) is learned through direct supervision. (c) During inference, the current action serves as the initial condition, and subsequent actions are recursively generated by numerically solving the ODE defined by the learned Action Transition Field.
SR (%) ↑
NLDLJ ↓
NSPARC ↓
EE-MAJ ↓
Boundary Jump ↓
BDR (%)
Benchmark / Task
Base
Ours
Base
Ours
Base
Ours
Base
Ours
Base
Ours
LIBERO
Spatial
98.8
99.3
26.35
25.32
2.121
1.983
1.19
0.48
0.04314
0.04283
0.72
Object
98.2
99.8
25.87
24.46
2.104
2.010
1.14
0.39
0.04102
0.04040
1.51
Goal
98.0
98.8
26.72
25.86
2.138
1.962
1.23
0.53
0.04577
0.04547
0.66
Long
92.4
95.1
28.24
26.44
2.172
2.042
1.31
0.38
0.05213
0.05061
2.92
Table 1: Comparison of task success rates and trajectory quality between ODEAction and Base Policy on LIBERO and EBench. SR gains are expressed in percentage points (pp), while improvements in NLDLJ, NSPARC, and EE-MAJ are relative reductions; EE-MAJ is reported in units of 10−3 . The mean BDR is the arithmetic mean across evaluation units, and aggregate results are calculated from unrounded data.
Figure 3: Relationship between improvements in boundary continuity and gains in task performance. (a) Four LIBERO suites and six EBench tasks arranged in descending order of Boundary Discontinuity Reduction (BDR). (b) Success-rate gains (SR gain) of ODEAction relative to Base Policy, shown in the same evaluation-unit order. BDR and success-rate gains exhibit a strong positive correlation across the ten evaluation units.
Method
Clean ↑
Random ↑
π0 [ 7 ]
65.9
58.4
π0.5 [ 8 ]
82.7
76.8
ABot-M0 [ 49 ]
86.1
85.1
X-VLA [ 53 ]
72.9
72.8
LingBot-VLA [ 48 ]
88.6
86.7
Base Policy [ 50 ]
85.7
87.3
Table 2: Success rates (%) on all 50 RoboTwin tasks. Base Policy denotes StarVLA-OFT [ 14 ] .
SR (%) ↑
NLDLJ ↓
NSPARC ↓
EE-MAJ ↓
Boundary Jump ↓
Task
Base
+RTC
+Legato
Ours
Base
+RTC
+Legato
Ours
Base
+RTC
+Legato
Ours
Base
+RTC
+Legato
Ours
Base
+RTC
+Legato
Ours
Adjust Bottle
92
91
93
94
23.68
23.66
23.71
21.54
1.998
2.030
2.051
1.991
1.095
1.068
1.020
0.625
0.1720
0.1562
0.1616
0.1507
Handover Block
79
80
75
82
29.10
28.17
27.59
26.67
2.850
2.954
2.868
2.735
1.248
1.050
0.833
0.581
0.0677
0.0630
0.0612
0.0391
Open Laptop
85
82
78
87
25.53
25.15
25.30
22.73
2.023
2.009
1.996
1.848
1.037
0.987
1.046
0.449
0.0781
0.0681
0.0660
0.0264
Lift Pot
88
90
86
91
23.34
22.64
22.59
20.69
1.723
1.681
1.698
1.566
1.170
0.817
0.724
0.511
0.1904
0.1712
0.1725
0.1613
Place Can Basket
68
63
58
68
25.19
24.63
24.57
23.18
2.554
2.502
2.550
2.338
0.848
0.709
0.662
0.500
0.0742
0.0540
0.0497
0.0462
Table 3: Comparison of four methods on five RoboTwin tasks, using StarVLA- π as the Base Policy. EE-MAJ is reported in units of 10−3 .
Method
SR (%) ↑
E↓
E1↓
E2↓
E2/E1↓
w/o neighborhood
47.4
8.84
5.78
11.90
2.06
w/o correction
56.2
13.17
8.26
18.08
2.19
ODEAction
84.4
4.25
4.42
4.07
0.92
Table 4: Ablation of key ODEAction design choices. The w/o neighborhood variant sets σ=0 , and the w/o correction variant sets k=0 . E denotes the full-trajectory offline rollout error, while E1 and E2 denote errors over the first and second halves, respectively; all are reported in units of 10−4 . E2/E1 is the error accumulation ratio.
Figure 4: Real-robot experimental platform, tasks, and quantitative evaluation. (a) AgileX Cobot Magic 1.0 dual-arm robotic platform and visual sensor configuration. (b) Real-robot tasks. (c) Comparison of average success rates (SR) across the four tasks. (d) Comparison of average trajectory quality across the four tasks.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Representative execution sequences for real-robot tasks. Each row corresponds to one task, with frames showing key stages of execution from left to right.
Task
C
R
Task
C
R
Adjust Bottle
100
100
Place Can Basket
74
74
Beat Block Hammer
95
94
Place Cans Plastic Box
100
99
Blocks Ranking RGB
99
98
Place Container Plate
99
99
Blocks Ranking Size
85
86
Place Dual Shoes
89
88
Click Alarm Clock
73
77
Place Empty Cup
97
97
Click Bell
84
87
Place Fan
91
92
Appendix
Table S1 : Per-task success rates (%) of Base Policy + ODEAction on all 50 RoboTwin tasks. C and R denote the Clean and Random evaluation settings, respectively.
Task
SR (%) ↑
NLDLJ ↓
NSPARC ↓
EE-MAJ ↓
Boundary Jump ↓
Put Bananas
70.00 / 60.00 / 70.00 / 75.00
25.84 / 25.38 / 25.17 / 24.66
2.757 / 2.691 / 2.656 / 2.595
2.62 / 2.49 / 2.42 / 1.79
0.1562 / 0.1519 / 0.1532 / 0.1474
Open Drawer
95.00 / 95.00 / 90.00 / 100.00
24.98 / 24.17 / 24.34 / 22.47
2.873 / 2.814 / 2.829 / 2.734
3.24 / 2.87 / 2.95 / 2.33
0.1147 / 0.1108 / 0.1089 / 0.1027
Place Cup
70.00 / 65.00 / 65.00 / 90.00
24.86 / 24.25 / 23.99 / 23.01
2.365 / 2.341 / 2.318 / 2.296
3.06 / 2.77 / 2.56 / 1.42
0.1517 / 0.1434 / 0.1411 / 0.1374
Insert Spoon
25.00 / 35.00 / 40.00 / 55.00
25.03 / 24.74 / 24.77 / 23.92
2.146 / 2.125 / 2.130 / 2.097
3.13 / 2.74 / 2.69 / 1.68
0.1455 / 0.1406 / 0.1419 / 0.1365
Average
65.00 / 63.75 / 66.25 / 80.00
25.18 / 24.64 / 24.57 / 23.52
2.535 / 2.493 / 2.483 / 2.431
3.01 / 2.72 / 2.66 / 1.81
0.1420 / 0.1367 / 0.1363 / 0.1310
Appendix
Table S2 : Comparison of success rates and control-command trajectory quality for four methods on four real-robot tasks. With π0.5 as the base policy, values in each cell correspond to Base Policy / +RTC / +Legato / +ODEAction , in that order. Higher SR is better, while lower values are better for the other four metrics; EE-MAJ is reported in units of 10−3 , and bold indicates the best results. Success rates are calculated from 20 trials per method per task, while trajectory-quality metrics are computed only from the successful episodes of each method.
Task
SR (%) ↑
NLDLJ ↓
NSPARC ↓
EE-MAJ ↓
Boundary Jump ↓
Adjust Bottle
92.0 / 94.0
32.87 / 21.54
3.891 / 1.991
5.614 / 0.625
0.2914 / 0.1507
Handover Block
67.0 / 82.0
37.26 / 26.67
4.622 / 2.735
8.327 / 0.581
0.2731 / 0.0391
Open Laptop
83.0 / 87.0
34.95 / 22.73
4.107 / 1.848
6.923 / 0.449
0.3236 / 0.0264
Lift Pot
88.0 / 91.0
33.58 / 20.69
3.673 / 1.566
8.541 / 0.511
0.4338 / 0.1613
Place Can Basket
62.0 / 68.0
35.76 / 23.18
4.414 / 2.338
6.350 / 0.500
0.2995 / 0.0462
Average
78.4 / 84.4
34.88 / 22.96
4.141 / 2.096
7.151 / 0.533
0.3243 / 0.0847
Appendix
Table S3 : Task performance and control-command trajectory quality of Direct Next-Action and ODEAction on five representative RoboTwin tasks. Values in each cell are ordered as Direct Next-Action / ODEAction , and bold indicates the best results. Higher SR is better, while lower values are better for the remaining metrics; EE-MAJ is reported in units of 10−3 .
Task
SR (%) ↑
NLDLJ ↓
NSPARC ↓
EE-MAJ ↓
Boundary Jump ↓
Adjust Bottle
91.0 / 94.0
22.31 / 21.54
2.004 / 1.991
0.813 / 0.625
0.1585 / 0.1507
Handover Block
74.0 / 82.0
28.36 / 26.67
2.865 / 2.735
1.061 / 0.581
0.0594 / 0.0391
Open Laptop
86.0 / 87.0
24.91 / 22.73
1.955 / 1.848
0.923 / 0.449
0.0695 / 0.0264
Lift Pot
89.0 / 91.0
21.78 / 20.69
1.638 / 1.566
0.781 / 0.511
0.1824 / 0.1613
Place Can Basket
53.0 / 68.0
24.71 / 23.18
2.494 / 2.338
0.750 / 0.500
0.0655 / 0.0462
Average
78.6 / 84.4
24.41 / 22.96
2.191 / 2.096
0.866 / 0.533
0.1071 / 0.0847
Appendix
Table S4 : Task performance and control-command trajectory quality of Shrinking Construction and Uniform ODEAction on five representative RoboTwin tasks. Values in each cell are ordered as Shrinking / Uniform ODEAction , and bold indicates the best results. Higher SR is better, while lower values are better for the remaining metrics; EE-MAJ is reported in units of 10−3 . The four trajectory-quality metrics are computed from the successful episodes of each method.
Task
E↓
E1↓
E2↓
E2/E1↓
Adjust Bottle
5.55 / 2.90
4.40 / 3.14
6.69 / 2.65
1.52 / 0.84
Handover Block
8.64 / 4.68
7.05 / 5.03
10.22 / 4.32
1.45 / 0.86
Open Laptop
7.39 / 3.79
5.35 / 3.90
9.42 / 3.67
1.76 / 0.94
Lift Pot
5.66 / 3.15
4.60 / 3.38
6.72 / 2.91
1.46 / 0.86
Place Can Basket
10.38 / 6.73
7.95 / 6.65
12.80 / 6.80
1.61 / 1.02
Average
7.52 / 4.25
5.87 / 4.42
9.17 / 4.07
1.56 / 0.92
Appendix
Table S5 : Offline recursive errors of Shrinking Construction and Uniform ODEAction on five representative RoboTwin tasks. Values in each cell are ordered as Shrinking / Uniform ODEAction , and bold indicates the best results. E , E1 , and E2 are reported in units of 10−4 , with lower values being better for all three; a lower E2/E1 indicates weaker relative error growth in the second half. The ratio in the Average row is obtained by dividing the cross-task average of E2 by the cross-task average of E1 .