NaP-Control: Navigating Diffusion Prior for Versatile and Fast Character Control

Chia-Wen Chen; Yan Wu; Korrawe Karunratanakul; Siyu Tang

cs.GRApr 15, 2026

NaP-Control: Navigating Diffusion Prior for Versatile and Fast Character Control

Authors: Chia-Wen Chen, Yan Wu, Korrawe Karunratanakul, Siyu Tang

Organizations: ETH Zurich

Abstract

Achieving precise, versatile whole-body character control in physics-based animation remains challenging. Recent diffusion-based policies generate rich and expressive motions but typically rely on gradient-based test-time guidance to satisfy task objectives, which is slow and can reduce robustness. We introduce NaP-Control (Navigating Diffusion Prior for Versatile and Fast Character Control), abbreviated as NaP. Our method uses reinforcement learning to manipulate the latent noise of a task-agnostic diffusion policy prior, steering it toward task-specific behaviors for fast, robust control with high motion fidelity. In contrast to methods that rely solely on offline training, NaP interacts with the environment during training to correct motions and optimize task rewards, improving success rates and enabling adaptation to challenging scenarios. By directly predicting task-optimized diffusion noise, NaP eliminates iterative guidance during denoising and enables efficient inference. Experiments show that NaP attains higher success rates and faster inference while preserving natural motion across diverse tasks.

Figures & tables

Figure 2: Framework overview. (a) The RL policy $\pi_{\theta}$ receives environment and proprioceptive states from the simulation. (b) The actor learns to predict optimal noise $\boldsymbol{\omega}\in\mathcal{W}$ aligned with task goals. (c) These predicted noises are denoised and decoded into executable actions $\mathbf{a}$ via a pretrained diffusion prior and a latent action decoder. Resulting transitions are then used to iteratively optimize the noise navigation policy with RL. (d) As a natural motion manifold, the diffusion prior can map predicted noises to future states $\mathbf{s}^{\text{prop}}$ and latent actions $\mathbf{z}$ conditioned on the trajectory history $\mathcal{H}$ .

Type	Methods	Jerk $(m/s^{3})\downarrow$	Success Rate $\uparrow$	Samples $\downarrow$
Flat	AMP [ 32 ]	2387	0.967	1200M
	CML [ 57 ]	2185	0.963	400M
	NaP(Ours)	1266	0.984	327M
Terrain	AdaptNet [ 58 ]	2444	0.868	408M
Terrain	NaP(Ours)	1788	0.860	900M

Table 1: Comparison with task-specific RL baselines on the far-goal reaching task.

Inference Speed (FPS) $\uparrow$	Far Goal Reaching	Velocity Control
UniPhys	2.9	5.2
NaP-Control (Ours)	22.5	24.4

Table 2: Inference efficiency comparison with diffusion-based baseline.

Methods	Far Goal Reaching		Agile Hand Reaching		Velocity Control			Object Interaction
Methods	Jerk $\downarrow$ $(m/s^{3})$	Success Rate $\uparrow$	Jerk $\downarrow$ $(m/s^{3})$	Success Rate $\uparrow$	Jerk $\downarrow$ $(m/s^{3})$	Velocity Error $\downarrow$ $(m/s)$	Success Rate $\uparrow$	Jerk $\downarrow$ $(m/s^{3})$	Success Rate $\uparrow$
PULSE [ 26 ]	2594	0.998	3631	0.998	2273	0.0455	0.996	1373	0.999
MaskedMimic [ 41 ]	3564	0.304	2750	0.725	2420	0.0678	0.355	1310	0.298
CLoSD [ 44 ]	1876	0.998	–	–	–	–	–	1505	0.862
UniPhys [ 53 ]	528	0.819	–	–	401	0.7971	0.921	–	–
NaP-Control(Ours)	1266	0.984	1146	0.991	752	0.1753	0.990	491	0.979

Table 3: Flat-ground motion naturalness and success rate comparison.

Figure 3: Qualitative Comparison of Object Interaction Task.

Methods	Far Goal Reaching		Agile Hand Reaching		Velocity Control
Methods	Jerk $(m/s^{3})$ $\downarrow$	Success Rate $\uparrow$	Jerk $(m/s^{3})$ $\downarrow$	Success Rate $\uparrow$	Jerk $(m/s^{3})$ $\downarrow$	Velocity Error $(m/s)$ $\downarrow$	Success Rate $\uparrow$
PULSE [ 26 ]	2830	0.966	3976	0.980	2397	0.0569	0.995
MaskedMimic [ 41 ]	3566	0.319	2723	0.676	2262	0.0556	0.292
NaP-Control(Ours)	1788	0.860	1520	0.909	1358	0.1654	0.950

Table 4: Uneven-terrain motion naturalness and success rate comparison. While MaskedMimic minimizes velocity error, it suffers from a lower success rate.

Panel of Figure 4 — Figure 4: Ablation studies. (a) Effect of state representation on flat-ground far goal reaching. (b–c) Effect of action chunk size $k$ for agile hand reaching on flat ground (b) and uneven terrain (c). (d) Comparison of joint state-action noise optimizing versus action-only noise optimizing for agile hand reaching.

Figure 5: Training convergence of the velocity control task on flat ground with different diffusion denoising steps.

Task	Task-specific input
Far goal reaching	Goal coordinates; terrain height maps (optional)
Agile hand reaching	Goal coordinates; terrain height maps (optional)
Velocity control	Target speed and direction; terrain height maps (optional)
Object interaction	Goal coordinates; sofa states
Shared input	Character’s proprioception

Table 5: Policy inputs for each task.

Symbol	Description	Value / range
Far goal reaching
$p_{\text{goal}}$	Target goal position in world coordinates.
$p_{\text{joint}}$	Pelvis position in world coordinates.
$\mathbf{d}_{\text{fwd}}$	Character forward heading unit vector.
$\mathbf{d}_{\text{goal}}$	Unit vector pointing from $p_{\text{joint}}$ to $p_{\text{goal}}$ .
$\alpha_{1},\alpha_{2}$	Scaling factors for $R_{\text{location}}$ and $R_{\text{orientation}}$ .	4.0 / 2.0 (uneven), 2.0

Table 6: Task-specific symbols used in task setup and reward design.

Category	Value / Configuration
Training Infrastructure
Hardware	RTX 4090 GPU, 4 CPUs, 32GB RAM
∗ Hardware	NVIDIA H200, 141GB RAM
Parallel Environments	256 / 1536 ∗
Inference Infrastructure	RTX TITAN GPU
PPO Hyperparameters

Table 7: Implementation details of the latent noise navigation policy. ∗ For velocity control task on uneven terrains.

Feature	Full State	Compact State	Root State Only
Global Root Trajectories	✓	✓	✓
6D Joint Rotations	✓	✓	–
Joint Angular Velocities	✓	✓	–
Local Joint Positions	✓	–	–
Local Joint Velocities	✓	–	–
Total Dimensions	398	256	47

Table 8: Comparison of state representation variants for the diffusion prior.

Methods	Jerk $(m/s^{3})\downarrow$	Success Rate $\uparrow$	Precision $(m)\downarrow$	FPS $\uparrow$
NaP(Ours)	570	0.972	0.186	34.3
TokenHSI [ 29 ]	575	0.995	0.089	53.9

Table 9: Comparison with TokenHSI on the path-following task. We report motion jerk, success rate, trajectory precision, inference speed, and training samples.

Diffusion Policy

Diffusion policies generate multimodal action sequences by iteratively denoising noise, enabling imitation, planning, and control in robotics.

See all 205 papers

GPC: Large-Scale Generative Pretraining for Transferable Motor Control
Jun 28, 2026Yi Shi, Yifeng Jiang, Chen Tessler +1RL Control Human Motion Generation
Simon Fraser University, Canada and NVIDIA, USA · NVIDIA, USA · Simon Fraser University, Canada and NVIDIA, Canada
Neural Assistive Impulses: Synthesizing Exaggerated Motions for Physics-based Characters
Apr 7, 2026Zhiquan Wang, Bedrich BenesPhysics-Based Simulation RL Control
Purdue University
FloodDiffusion 2: Efficient and Path Controllable Streaming Motion Generation
Sep 27, 2026Yiyi Cai, Yuhan Wu, Kunhang Li +6Motion Diffusion Models Human Motion Generation
Alaya Lab · The University of Tokyo · Japan Advanced Institute of Science and Technology +1

Jun 28, 2026J/K move · Enter open · S save

GPC: Large-Scale Generative Pretraining for Transferable Motor Control

Yi Shi, Yifeng Jiang, Chen Tessler, Xue Bin Peng

Simon Fraser University, Canada and NVIDIA, USA · NVIDIA, USA · Simon Fraser University, Canada and NVIDIA, Canada

RL Control Human Motion Generation AI Control Transfer Learning

Open paper AskPDF arXiv

Developing controllers capable of completing a wide range of tasks in a natural and life-like manner is a key challenge in enabling practical applications of physics-based character animation. In this work, we introduce Generative Pretrained Controllers (GPC), which leverage tokenization and next-token modeling to create general-purpose, reusable generative controllers from large-scale motion datasets. Our framework utilizes end-to-end reinforcement learning to jointly optimize a "motion vocabulary", modeled via Finite Scalar Quantization (FSQ), along with a corresponding control policy that can map the discrete codes to physics-based controls. After the "codebook" has been learned, the underlying structure of this large vocabulary is modeled by training a GPT-style autoregressive transformer, leading to a powerful generative controller that generates controls for a physically simulated character by performing next-token prediction. Once the generative controller has been trained, we propose a suite of adaptation techniques for finetuning the controller for new downstream tasks. Our proposed framework greatly simplifies the training process compared to previous tokenized methods, and achieves a 99.98% success rate in reproducing a vast corpus of motion clips. The generative controller exhibits a variety of natural emergent behaviors, such as responsive behaviors to perturbations and recovery behaviors after falling. This results in highly robust general purpose controllers for a variety of downstream applications.

Abstract

Diffusion Policy

Explore similar work

GPC: Large-Scale Generative Pretraining for Transferable Motor Control

Neural Assistive Impulses: Synthesizing Exaggerated Motions for Physics-based Characters

FloodDiffusion 2: Efficient and Path Controllable Streaming Motion Generation