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

Zhiquan Wang; Bedrich Benes

cs.AIApr 7, 2026

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

Authors: Zhiquan Wang, Bedrich Benes

Organizations: Purdue University

Abstract

Physics-based character animation has become a fundamental approach for synthesizing realistic, physically plausible motions. While current data-driven deep reinforcement learning (DRL) methods can synthesize complex skills, they struggle to reproduce exaggerated, stylized motions, such as instantaneous dashes or mid-air trajectory changes, which are required in animation but violate standard physical laws. The primary limitation stems from modeling the character as an underactuated floating-base system, in which internal joint torques and momentum conservation strictly govern motion. Direct attempts to enforce such motions via external wrenches often lead to training instability, as velocity discontinuities produce sparse, high-magnitude force spikes that prevent policy convergence. We propose Assistive Impulse Neural Control, a framework that reformulates external assistance in impulse space rather than force space to ensure numerical stability. We decompose the assistive signal into an analytic high-frequency component derived from Inverse Dynamics and a learned low-frequency residual correction, governed by a hybrid neural policy. We demonstrate that our method enables robust tracking of highly agile, dynamically infeasible maneuvers that were previously intractable for physics-based methods.

Figures & tables

Figure 1: Overview of the Hybrid Dynamics Architecture: We decouple the control problem into two parallel streams. The Analytical Stream (Top) serves as an open-loop feed-forward guide, employing an RNEA solver to derive a nominal Impulse Reference from the target motion. The Neural Stream (Bottom) operates as a closed-loop feedback controller; the Control Policy observes the current simulation State and the target reference to predict a Residual Impulse , modulated by a learned Gate . These components are dynamically fused to drive the Physics Simulation , enabling the character to robustly track exaggerated maneuvers that are intractable for purely analytical or purely learning-based methods.

Figure 2: Comparison between Force-space and Momentum-space control signals. We show the magnitude profiles of the assistive intervention for a simulated character performing the motion described in the teaser. While the instantaneous Assistive Force (solid lines) suffers from extreme magnitude spikes and dependency on simulation time-steps ( $F\propto 1/\Delta t$ ), the integrated Assistive Impulse (blue dashed) provides a smooth, bounded, and frequency-invariant learning target for the neural policy.

Figure 3: The detailed architecture of the proposed Neural Assistive Impulse (NAI) policy network.

Figure 4: Snapshots of the simulated humanoid characters trained using our NAI framework, showcasing exaggerated and stylized motion capabilities.

Figure 5: Snapshots of the simulated humanoid characters trained using our NAI framework, showcasing exaggerated and stylized motion capabilities.

Motion Sequence	ADD/AMP (Baseline)	NAI (ours)
Dashing Attack	0%	100%
Ground Dashing	0%	100%
Aerial Dashing	0%	100%
Multi-Dir Combat	0%	100%
Aerial Launcher	0%	100%
Gravity-Defying	0%	100%

Table 1: Comparison of Tracking Success Rate Across Exaggerated Motion Categories. The standard underactuated baseline (ADD/AMP) exhibits a 0% success rate across all test sequences, as pure joint torques are mathematically insufficient to execute the required momentum discontinuities. Our proposed NAI framework achieves a 100% success rate across the identical test set.

Task	Method	Pos Error [m]		Vel Error [rad/s]		Total Imp. $\downarrow$		Ref Imp.		Res Imp.		Jitter $\downarrow$
		Mean	Std	Mean	Std	Lin	Ang	Lin	Ang	Lin	Ang	[Unit]
Ground Dashing	NAI	0.022	0.032	0.42	0.064	15.92	4.93	5.51	3.82	11.53	2.18	8.7
	Naive	0.007	0.002	0.161	0.060	20.64	2.43	N/A	N/A	N/A	N/A	14.7
Dashing Punch	NAI	0.020	0.039	0.48	0.112	34.93	8.84	21.50	7.60	16.92	2.72	8.84
	Naive	0.031	0.031	0.277	0.116	36.54	2.43	N/A	N/A	N/A	N/A	10.91
Aerial Dashing	NAI	0.013	0.006	0.17	0.070	21.46	2.30	22.56	3.13	0.86	2.27	3.6

Table 2: Quantitative comparison against the baseline. We report Position (Pos) and Velocity (Vel) errors, followed by Impulse analysis. Each task compares our Dual-Gated method against the Naive Baseline . (N/A indicates unavailable baseline data). All the motion skills are evaluated with 4096

Figure 6: Comparison of assistive intervention profiles of teaser motion. The Native Assist (red) exhibits continuous, high-frequency force fluctuations ("always-on"), indicating an over-reliance on external assistance to force kinematic compliance. In contrast, NAI (blue) demonstrates distinct sparsity : the assistive impulse drops to near-zero during physically consistent phases (e.g., $t=0.0s-0.3s$ , $t=3.5s+$ ) and activates smoothly only when dynamic transients require momentum injection.

Figure 7: Tracking error comparison without external perturbations. The Open-Loop Offline RNEA (blue curve) accumulates numerical integration drift monotonically over time, leading to state divergence. The proposed closed-loop policy (orange curve) maintains bounded tracking errors.

Figure 8: The linear ( $\beta_{lin}$ ) and angular ( $\beta_{ang}$ ) gating scalars. Both parameters are dynamically modulated within the $[0.3,0.6]$ interval around a mean value of $0.4$ . This indicates a consistent allocation of approximately $60\%$ of the assistive intervention to the neural residual.

Figure 9: Tracking error comparison under external perturbation. Upon physical impact. The proposed closed-loop policy dampens the applied external impulse and synthesizes corrective actions to converge back to the reference trajectory.

Figure 10: Ablation analysis evaluating the impact of the Shadow Compass and Sparsity loss formulations on the training success rate. The full NAI framework (blue curve) exhibits the most rapid convergence. Removing the Shadow Compass Loss (green curve) severely delays convergence due to inefficient exploration in the unguided directional space. The NAI - Baseline (brown curve), lacking both regularizations, exhibits the most severe optimization failure.

Figure 11: Ablation analysis evaluating the impact of the Shadow Compass and Sparsity loss formulations on the mean body position error. Removing the Sparsity Loss (red curve) introduces numerical drift from continuous unconstrained force injections, resulting in suboptimal error metrics during the mid-to-late training iterations. The NAI - Baseline (brown curve) consistently fails to minimize the tracking error effectively.

Configuration	Control Jitter
NAI - Baseline	3.45
NAI - No Sparsity	2.14
NAI - No Compass	2.04
Full NAI (Ours)	2.29

Table 3: Quantitative comparison of control signal jitter across ablation configurations. The Sparsity formulation eliminates the oscillations of the baseline while preserving the necessary high-frequency impulse spikes (yielding a nominally higher jitter than the over-damped, non-physical No-Sparsity configuration).

Physics-Based Simulation

Physics-based simulation models physical systems with equations and numerical solvers, increasingly augmented by learned models for speed and control.

See all 228 papers

Apr 15, 2026

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

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.

Chia-Wen Chen, Yan Wu, Korrawe Karunratanakul +1

ETH Zurich

Diffusion Policy Diffusion Models for Motion Generation RL for Diffusion Models

Ask PDF

Sep 9, 2026

InstantMimic: A High Performance System for Learning Physics-based Skills in Seconds

Physics-based character control is a long-standing challenge in computer graphics and robotics, requiring policies that satisfy complex dynamics while producing realistic motion. Recent Deep RL approaches, particularly imitation learning methods such as DeepMimic, have had broad impact beyond animation, influencing robotics by enabling agile and expressive behaviors. While these approaches achieve impressive results, they remain computationally inefficient to train in practice. Despite GPU-accelerated simulation, we find that end-to-end pipelines often underutilize hardware due to overheads outside the physics solver, caused by fragmented GPU kernels and CPU memory access in the critical path. We present InstantMimic, a system that addresses these inefficiencies by making the entire training loop GPU-native. Built on a GPU-native physics backend, our unified pipeline integrates simulation, environment computation, policy inference, and policy updates within a single execution flow. As a result, InstantMimic reduces training time for diverse physics-based skills to a few seconds and makes LLM-agent-driven hyperparameter search practical.

Ikjun Choi, Geonho Leem, Jungdam Won

Seoul National University, Seoul, South Korea

Reinforcement Learning Imitation Learning GPU-Accelerated RL Physics-Based Simulation

Ask PDF

Mar 15, 2026

Physics-Informed Policy Optimization via Analytic Dynamics Regularization

Reinforcement learning (RL) has achieved strong performance in robotic control; however, state-of-the-art policy learning methods, such as actor-critic methods, still suffer from high sample complexity and often produce physically inconsistent actions. This limitation stems from neural policies implicitly rediscovering complex physics from data alone, despite accurate dynamics models being readily available in simulators. In this paper, we introduce a novel physics-informed RL framework, called PIPER, that seamlessly integrates physical constraints directly into neural policy optimization with analytical soft physics constraints. At the core of our method is the integration of a differentiable Lagrangian residual as a regularization term within the actor's objective. This residual, extracted from a robot's simulator description, subtly biases policy updates towards dynamically consistent solutions. Crucially, this physics integration is realized through an additional loss term during policy optimization, requiring no alterations to existing simulators or core RL algorithms. Extensive experiments demonstrate that our method significantly improves learning efficiency, stability, and control accuracy, establishing a new paradigm for efficient and physically consistent robotic control.

Namai Chandra, Liu Mohan, Zhihao Gu +1

Electronic Systems, IIT Madras, India · EmPACT Lab, Nanyang Technological University, Singapore

Physics-Informed RL RL Control Robotic Control

Ask PDF

Abstract

Physics-Based Simulation

Explore similar work

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

InstantMimic: A High Performance System for Learning Physics-based Skills in Seconds

Physics-Informed Policy Optimization via Analytic Dynamics Regularization