Physics-based humanoid control relies on training with motion datasets that have diverse data distributions. However, the fixed difficulty distribution of datasets limits the performance ceiling of the trained control policies. Additionally, the method of acquiring high-quality data through professional motion capture systems is constrained by costs, making it difficult to achieve large-scale scalability. To address these issues, we propose a closed-loop automated motion data generation and iterative framework. It can generate high-quality motion data with rich action semantics, including martial arts, dance, combat, sports, gymnastics, and more. Furthermore, our framework enables difficulty iteration of policies and data through physical metrics and objective evaluations, allowing the trained tracker to break through its original difficulty limits. On the PHC single-primitive tracker, using only approximately 1/10 of the AMASS dataset size, the average failure rate on the test set (2201 clips) is reduced by 45% compared to the baseline. Finally, we conduct comprehensive ablation and comparative experiments to highlight the rationality and advantages of our framework.
Figures & tables
Figure 1: Overview of the CLAIMS pipeline: a closed-loop system that refines prompts from a 5-domain library (martial arts, dance, combat, sports, gymnastics), synthesizes motions with MDM, filters them by physics and VLM checks, trains humanoid trackers with RL, and uses multimodal feedback to generate progressively harder tasks.
Figure 2: Difficulty-aware variable library across five domains and four compositional axes.
Figure 3: Prompt-to-prompt data generation.
Figure 4: Competitive iteration between the controller and the dataset
Figure 5: Schematic Diagram of the Automated Iterative Loop
Figure 6: Loop-wise Difficulty and Speed Trends of Qwen Evaluations
Figure 7: (a) The t-SNE graph of martial arts datasets comparison; (b) The t-SNE graph of all datasets(AMASS/Ours/Random Prompts/HumanAct).
Figure 8: Visualization of 5 categories of professional movements
Method
Kungfu
EMDB
AIST++
VC
Avg
AMASS
47.1
53.3
67.6
31.2
58.3
L0
37.8
31.1
68.8
33.3
55.9
L1
47.7
33.3
75.3
38.7
64.0
L2
51.7
51.1
73.3
41.6
63.8
L3
59.1
64.4
82.1
50.9
72.4
L4
55.8
55.6
84.0
45.1
71.8
Table 1: Success rate (%) of different pipelines across test sets (totaling 2201 clips). The test suite includes: Motion-X/Kungfu (663), EMDB (45), AIST++ (1320), and Video-Convert (173). We report per-set results and the average (Avg) calculated by clip.
Dataset
AMASS
loop0
loop1
Motion-X/Kungfu
57.2
54.0
65.8
EMDB
53.3
48.9
71.1
AIST++
68.9
75.3
83.9
Video-Convert
41.6
47.4
62.4
Table 2: Maskedmimic FC*: Success Rate on each test set.”*” means early period.
Figure 9: Qualitative tracking performance between Ours and Baselines of PHC and Maskedmimic.
Method
Motion-X/Kungfu
EMDB
AIST++
Video-Convert
L1
L2
L3
L1
L2
L3
L1
L2
L3
L1
L2
L3
full (var)
47.7
51.7
59.1
33.3
51.1
64.4
75.3
73.3
82.1
38.7
41.6
50.9
w/o var
53.5
59.1
59.9
53.3
57.8
53.3
75.6
80.2
76.5
50.3
48.0
46.8
w/o VLMs
46.3
54.8
58.4
42.2
62.2
60.0
73.8
80.0
81.9
43.4
45.1
57.2
w/o metrics
48.0
50.7
54.6
31.1
31.1
35.6
78.3
77.5
78.9
38.2
28.9
40.5
w/o VLMs+metrics
46.5
54.3
54.3
35.6
46.7
46.7
71.0
76.1
77.4
38.2
46.8
50.9
Table 3: Ablation study on success rate across different test sets and loops.
Dataset
Loop0(1400 clips)
Loop6
SR
SR
Kungfu(663)
58.7
60.3
EMDB(45)
73.3
64.4
AIST++(1320)
85.3
88.1
Video-Convert(173)
55.5
59.0
Table 4: Metrics(Success Rate) on each dataset for 1400 clips testset Loop0 and our Loop6.
Figure 10: The velocity distribution of AMASS and ours.
Data Loop
SR
g-MPJPE
Acc
Vel
L0
75.3
49.78
5.97
8.54
L1
65.8
53.84
6.66
9.34
L2
65.2
61.29
7.65
10.86
L3
61.2
57.24
7.03
9.95
L4
59.0
57.70
7.08
9.99
L5
52.7
59.10
7.49
10.65
Table 5: PHC+ tracking performance on our datasets.A lower success rate indicates that the motions are more challenging for third-party trackers.
State Key Laboratory of General Artificial Intelligence, Beijing Institute for General Artificial Intelligence (BIGAI) · Joint Laboratory of Embodied AI and Humanoid Robots, BIGAI & Unitree Robotics · Shanghai Jiao Tong University +4