StepControl: Long-Horizon Human Motion Generation from Foot-Contact Schedules
Authors: Hojun Jang, Jinseok Bae, Donggeun Lim, Taesun Kim, Seungbin Jeong, Younghwan Lee, Yong Seok Kim, Young Min Kim
Organizations: Dept. of Electrical and Computer Engineering, Seoul National University · Interdisciplinary Program in Artificial Intelligence, Seoul National University · Institute of New Media and Communications, Seoul National University
Controllable human motion generation has largely focused on specifying what action should occur or where the body should move. We study a complementary question: can full-body motion be controlled by specifying only when each foot contacts the ground? We introduce StepControl, an autoregressive diffusion model conditioned on a future binary left/right foot-contact schedule, without requiring target root and body poses, or semantic prompts. StepControl represents contact as explicit temporal events and injects them through dedicated cross-attention with classifier-free guidance. We further introduce a stochastic mixture noise schedule and warmup strategy for stable long-horizon autoregressive generation. Experiments show that StepControl follows prescribed contact timing more accurately than adapting existing spatially guided motion generators to contact constraints. It maintains effective contact control over long continuous rollouts and responds to edited schedules such as cadence changes, stop-and-go pauses, limping gaits, and bilateral jumping. Our results show that minimal foot-contact schedules provide an effective and intuitive temporal interface for controllable motion generation.
Figures & tables
Figure 1: StepControl generates full-body motion from foot-contact schedules without future foot positions or root trajectories. Beyond reproducing ordinary walking cadence, StepControl follows atypical asymmetric schedules.
Figure 2: Overview of StepControl. (a) Motion tokens with per-frame diffusion levels and frame embeddings attend to event condition tokens via dedicated per-layer cross-attention, predicting clean motion X^(0) for continuous long-horizon generation. (b) Within each sliding window, 2D foot-contact signals C are expanded into 9D event-aware representations E . Conditioned on E , the denoiser steps down noisy future latents via per-frame DDIM updates ( ki→ki′ ), commits the earliest finalized frame x^t(0) to clean history, and appends a new terminal noise state ( k=99 ) as the window slides forward.
Rollout Horizon
Method
Bal. Acc. (%) ↑
Tr. F1 (%) ↑
Slip (mm/f) ↓
RA-LB (mm) ↓
Mean Pen. (mm) ↓
1-Second
DART
74.34
15.26
7.84
109.97
0.30
ARDY
69.37
16.92
20.69
158.62
0.47
StepControl
79.90
36.15
17.63
144.25
0.00
StepControl (+Opt.)
80.60
33.59
8.64
144.25
0.00
5-Second
DART
69.25
16.52
3.28
180.68
0.31
ARDY
64.28
10.29
13.52
185.86
0.34
Table 1: Quantitative comparison on autoregressive motion generation across extended horizons. We evaluate contact synchronization and progressive rollout stability against autoregressive baselines (DART [ 38 ] , ARDY [ 39 ] ) across LAFAN1 [ 10 ] test sequences.
Figure 3: 20 s Motion Generation Result of StepControl. StepControl generates natural and consistent long-horizon motion from only foot-contact schedule.
Figure 4: Qualitative comparison on contact schedule editing (bilateral jump). StepControl better synchronizes simultaneous flight and landing phases, whereas DART [ 38 ] fails to coordinate flight and ARDY [ 39 ] exhibits noisy contact timing and unstable poses. ARDY is rendered with its native body model.
Modification
Method
Bal. Acc. ↑ (%)
Tr. F1 ↑ (%)
Slip ↓ (mm/f)
Cadence Scaling
DART
67.91
9.13
3.44
ARDY
77.41
16.44
6.19
StepControl
76.11
33.57
16.49
Stop-and-Go
DART
69.49
21.13
2.80
ARDY
78.88
16.81
4.05
StepControl
82.18
37.31
11.77
Table 2: Quantitative evaluation of interactive contact scheduling (5 s horizon). We evaluate responsiveness to temporal cadence scaling, gait transitions, and atypical gait synthesis against DART [ 38 ] and ARDY [ 39 ] across LAFAN1 [ 10 ] test sequences.
Ablation Configuration
Bal. Acc. (%) ↑
Tr. F1 (%) ↑
Slip (mm/f) ↓
RA-LB (mm) ↓
Bo. Jerk (mm/f 3 ) ↓
(a) Contact Feature Representation
Raw 2D Contact
76.23
21.79
16.06
182.54
77.92
Delta 5D Contact
79.30
23.79
17.45
175.75
75.48
Event 9D Contact (Ours)
79.30
39.58
15.54
181.64
70.54
(b) Training Dataset
LAFAN1 Only
80.99
26.00
21.02
180.43
83.34
Table 3: Ablation studies on the LAFAN1 [ 10 ] test benchmark. We evaluate 5-second prefixes of the same 20-second evaluation segments used in the main benchmark.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Per-Frame
20 s Rollout
Throughput
DART [ 38 ] (50-step Opt.)
≈2267ms
≈907s
0.44fps
ARDY [ 39 ] (10-step Diff.)
≈75.0ms
≈30.0s
13.3fps
StepControl (Ours)
7.5ms
2.9s
133.3fps
Appendix
Table A1 : Inference runtime and generation throughput comparison for 20-second (400-frame) rollouts on a single NVIDIA RTX 4090 GPU. StepControl achieves real-time throughput ( >130 fps) via direct single-pass generation.
Loss Formulation
Bal. Acc. (%) ↑
Tr. F1 (%) ↑
Slip (mm/f) ↓
RA-LB (mm) ↓
Bo. Jerk (mm/f 3 ) ↓
Only Motion Reconstruction Loss ( Lmse )
73.87
21.60
22.40
204.35
71.45
w/o Stance Velocity Loss ( Lstance )
76.84
32.27
19.93
201.03
70.94
w/o Foot Height Loss ( Lheight )
78.02
33.30
15.39
180.55
70.68
w/o 4D Contact MSE Loss ( Lcontact4 )
79.03
34.21
15.80
181.17
70.28
Full Multi-Task Loss (StepControl)
79.30
39.58
15.54
181.64
70.54
Appendix
Table A2 : Ablation on loss formulations (LAFAN1 test benchmark). All variants use the same mixture noise schedule, pretraining initialization, fine-tuning protocol, evaluation segments, and random seed.
CFG Guidance Scale
Bal. Acc. (%) ↑
Tr. F1 (%) ↑
Slip (mm/f) ↓
RA-LB (mm) ↓
Bo. Jerk (mm/f 3 ) ↓
s=0.0 (Unconditional Baseline)
52.08
3.90
38.36
208.98
69.12
s=1.0 (Standard Conditional Prior)
74.67
24.52
16.91
183.16
69.49
s=2.0 (Default StepControl)
79.30
39.58
15.54
181.64
70.54
Appendix
Table A3 : Ablation on classifier-free guidance (CFG) scale s (LAFAN1 test benchmark). Increasing s enhances temporal contact adherence and transition precision.
Inference
Bal. Acc. (%) ↑
Tr. F1 (%) ↑
Slip (mm/f) ↓
RA-LB (mm) ↓
Bo. Jerk (mm/f 3 ) ↓
Chunk AR
76.32
30.75
19.96
203.63
75.59
Progressive (Ours)
79.30
39.58
15.54
181.64
70.54
Appendix
Table A4 : Rollout strategy comparison on the LAFAN1 test benchmark.
Low
Mid
High
Transition rate (events/s)
1.2–3.2
3.4–4.0
4.4–8.0
Windows
20
16
16
Target events
253
300
460
DART
17.82
15.93
16.88
ARDY
13.42
16.64
18.54
Raw 2D Contact
26.55
33.34
25.65
Appendix
Table A5 : Transition F1 (%) by contact-transition rate.
Training-free controllable motion generation has attracted growing interest for enabling flexible constraint enforcement without constraint-specific training. However, existing training-free methods require constraints to be continuous objective-based with differentiable losses, while many real-world requirements are criterion-based and provide only discontinuous, sparse, or even black-box feedback. In this paper, we propose Motion-Inference-as-Control (MIC), the first training-free motion generation framework that handles both continuous objective-based and criterion-based motion constraints under a shared mechanism. The key idea is to cast diffusion-based motion generation as a stochastic control problem. This perspective not only provides principled and practically effective step-wise control laws that support criterion-based constraints without requiring differentiability and naturally accommodate objective-based constraints as a special case, but also motivates a control-oriented constraint coordination mechanism that adaptively balances and reconciles motion constraints during generation. Experiments across diverse constraint settings demonstrate the effectiveness of our framework.
Xiaofei Hui, Bo Yan, Haoxuan Qu +2
Lancaster University, United Kingdom · Shenzhen International Graduate School, Tsinghua University, China
Long-range human movement generation remains a central challenge in computer vision and graphics. Generating coherent transitions across semantically distinct motion domains remains largely unexplored. This capability is particularly important for applications such as dance choreography, where movements must fluidly transition across diverse stylistic and semantic motifs. We propose a simple and effective inference-time optimization framework inspired by diffusion-based stochastic optimal control. Specifically, a control-energy objective that explicitly regularizes the transition trajectories of a pretrained diffusion model. We show that optimizing this objective at inference time yields transitions with fidelity and temporal coherence. This is the first work to provide a general framework for controlled long-range human motion generation with explicit transition modeling.
Haichao Wang, Alexander Okupnik, Yuxing Han +3
Shenzhen International Graduate School, Tsinghua University, China · Institute of Information Systems, Universität Liechtenstein, Liechtenstein · Department of Computer Science, New York University, USA +1
High-quality human motion data is becoming increasingly important for applications in robotics, simulation, and entertainment. Recent generative models offer a potential data source, enabling human motion synthesis through intuitive inputs like text prompts or kinematic constraints on poses. However, the small scale of public mocap datasets has limited the motion quality, control accuracy, and generalization of these models. In this work, we introduce Kimodo, an expressive and controllable kinematic motion diffusion model trained on 700 hours of optical motion capture data. Our model generates high-quality motions while being easily controlled through text and a comprehensive suite of kinematic constraints including full-body keyframes, sparse joint positions/rotations, 2D waypoints, and dense 2D paths. This is enabled through a carefully designed motion representation and two-stage denoiser architecture that decomposes root and body prediction to minimize motion artifacts while allowing for flexible constraint conditioning. Experiments on the large-scale mocap dataset justify key design decisions and analyze how the scaling of dataset size and model size affect performance.