CLASH: Collision Learning via Augmented Sim-to-real Hybridization to Bridge the Reality Gap

Haotian He; Ning Guo; Siqi Shi; Qipeng Liu; Wenzhao Lian

cs.ROFeb 21, 2026

CLASH: Collision Learning via Augmented Sim-to-real Hybridization to Bridge the Reality Gap

Authors: Haotian He, Ning Guo, Siqi Shi, Qipeng Liu, Wenzhao Lian

Organizations: School of Mathematical Sciences, Peking University. · School of Artificial Intelligence, Shanghai Jiao Tong University. · School of Mathematical Sciences, Beijing Normal University.

Abstract

The sim-to-real gap, particularly in the inaccurate modeling of contact-rich dynamics like collisions, remains a primary obstacle to deploying robot policies trained in simulation. Conventional physics engines often trade accuracy for computational speed, leading to discrepancies that prevent direct policy transfer. To address this, we introduce Collision Learning via Augmented Sim-to-real Hybridization (CLASH), a data-efficient framework that learns a parameter-conditioned impulsive collision surrogate model and integrates it as a plug-in module within a standard simulator. CLASH first distills a base model from an imperfect simulator (MuJoCo) using large-scale simulated collisions to capture reusable physical priors. Given only a handful of real collisions (e.g., 10 samples), it then (i) performs gradient-based identification of key contact parameters and (ii) applies small-step, early-stopped fine-tuning to correct residual sim-to-real mismatches while avoiding overfitting. The resulting hybrid simulator not only achieves higher post-impact prediction accuracy but also reduces the wall-clock time of collision-heavy CMA-ES search by 42-48% compared to MuJoCo. We demonstrate that policies obtained with our hybrid simulator transfer more robustly to the real world, doubling the success rate in sequential pushing tasks with reinforcement learning and significantly increase the task performance with model-based control.

Figures & tables

Fig. 1 : Overview of our CLASH framework and learning downstream tasks with the hybrid simulator.

Method	Simulator Dependency	System Identification	Hybrid Model	Real Data Requirement	Accuracy
MuJoCo [ 2 ]	✓	✓(0th-order)	✗	Few	Low
MJX [ 21 ]	✓	✓(1st-order)	✗	Few	Low
Residual Learning [ 18 ]	✓	✗	✓	Few	High
NeuralSim [ 17 ]	✗	✗	✗	Many	High
CLASH(ours)	✓	✓(1st-order)	✓	Few	High

TABLE I : Comparison with Prior Sim-to-Real Modeling Methods

Fig. 2 : Our framework Collision Learning via Augmented Sim-to-real Hybridization (CLASH): First collect collision dataset from physics simulator and pre-train a base model to learn the collision dynamics, then adopt the model to identify the system parameters for real-world collision. Subsequently, fine-tune with real-world dataset to obtain a collision model. Finally, integrate the collision model with the physics simulator to build a hybrid simulator that is close to reality.

Panel of Figure 4 — Fig. 3 : (a) The proposed hybrid simulator is applied to two downstream tasks: model-based optimization and reinforcement learning. (b) System setup, where a Franka arm manipulates an impact block to strike the semi-cylinder block.

Fig. 4 : Collision model accuracy ( $\alpha=0.1$ ) in different settings. Our fine-tuned strategy effectively avoids overfitting on scarce real-world data and achieves higher accuracy than MuJoCo, whose performance is limited by unmodeled effects.

Symbol	Definition	Range
$v_{pre}$	Velocity before collision	[ $0.2,0.8$ ]
$\theta$	Deflection angle of the impact point relative to the center of the semicircle	[ $\frac{\pi}{12},\frac{11\pi}{12}$ ]
$\theta_{1}$	Deflection angle of the velocity relative to the impact point	[ $-\frac{\pi}{12},\frac{\pi}{12}$ ]
$\mu_{1}$	Friction coefficient between impact block and semi-cylinder	[ $0.05,1$ ]
$\mu_{2}$	Friction coefficient between ground and semi-cylinder	[ $0.02,1$ ]
$e_{1}$	Damping ratio between impact block and semi-cylinder	[ $0.1,3$ ]

TABLE II : Simulation Settings

Method	Semi-Cylinder	Square	Triangle
MuJoCo	0.64	0.50	0.49
NN(w/o Pre-train)	0.42	0.42	0.36
NN(ours)	0.76	0.55	0.54

TABLE III : Collision Model Accuracy(

\alpha=0.1

) For Three Shape Blocks.

Simulator	Pos/m	Ori/rad	Pos/m	Ori/rad	Pos/m	Ori/rad
	Semi-Cylinder		Square		Triangle
MuJoCo	0.0245( $\pm$ 0.0176)	0.7882( $\pm$ 0.6071)	0.0222( $\pm$ 0.0124)	1.2696( $\pm$ 0.6926)	0.0251( $\pm$ 0.0154)	0.6673( $\pm$ 0.4036)
Hybrid Simulator (w/o Pre-train)	0.0363( $\pm$ 0.0156)	0.5751( $\pm$ 0.5074)	0.0243( $\pm$ 0.0140)	1.9751( $\pm$ 0.8621)	0.0303( $\pm$ 0.0187)	0.7027( $\pm$ 0.5069)
Hybrid Simulator (Ours)	0.0159( $\pm$ 0.0046)	0.4152( $\pm$ 0.3014)	0.0163( $\pm$ 0.0080)	1.2479( $\pm$ 1.0630)	0.0185( $\pm$ 0.0092)	0.7319( $\pm$ 0.5518)

TABLE IV : Error in Model-Based Optimization

Time(s)	Semi-cylinder	Square	Triangle
MuJoCo	27.42( $\pm$ 3.24)	28.58( $\pm$ 3.71)	25.20( $\pm$ 2.89)
Hybrid Simulator (Ours)	15.81( $\pm$ 2.75)	14.74( $\pm$ 1.83)	13.21( $\pm$ 3.52)

TABLE V : Time Cost in CMA-ES

Fig. 5 : Linear and angular velocities of the square block in simulation and real world. MuJoCo captures the linear velocity decay reasonably well, but fails to reproduce the angular velocity decay accurately. Raw: raw measurements from the real world. Smoothed: filtered real-world data. Simulation: MuJoCo-simulated data.

Panel of Figure 11 — Fig. 6 : (a) The three block shapes used in the model-based optimization experiments: semi-cylinder, square, and triangle (from left to right). (b) The designated workspace, bounded by the two white lines. (c) The three evaluation routes used in the RL experiments. The difficulty increases from Route 1 to Route 3. For Route 1 and 2, the initial orientation of the semi-cylinder is randomized, while for Route 3 it is fixed.

Term	Formula	Description
$R_{1}$	$\exp(-\Delta x^{2})$	Dense position reward
$R_{2}$	$\frac{1+\cos(\Delta\theta)}{2}$	Orientation reward at sub-goal
$R_{3}$	$(T-t)$	Strike efficiency reward at sub-goal
$R_{4}$	$r_{0}$	Penalty for leaving safe zone

TABLE VI : Reward function design for policy learning. The total reward is

R=R_{1}+R_{2}+R_{3}+R_{4}

Simulator	SR	SR	SGCR	SR	SGCR
	Route 1	Route 2		Route 3
MuJoCo	4/10	2/10	8/20	1/10	13/30
Hybrid Simulator (ours)	8/10	5/10	14/20	5/10	21/30

TABLE VII : Results in Sequential Semi-Cylinder Pushing Task

Sim-to-Real Transfer

Simulation-to-reality transfer adapts models and policies trained in simulation for reliable performance in physical or otherwise real-world settings.

See all 315 papers

Few-Shot Neural Differentiable Simulator: Real-to-Sim Rigid-Contact Modeling
Mar 6, 2026Zhenhao Huang, Siyuan Luo, Bingyang Zhou +3Differentiable Physics Robot Dynamics Learning
National University of Singapore, Singapore · Prana Lab, San Francisco, USA
An Real-Sim-Real (RSR) Loop Framework for Generalizable Robotic Policy Transfer
Mar 13, 2025Yuxuan Xu, Shiyu Wang, Jinhao Huang +6Robotic Data Collection Sim-to-Real Transfer
Institute for AI Industry Research (AIR), Tsinghua University · Beijing Jiaotong University · The Hong Kong University of Science and Technology (Guangzhou)
ConCent: Contact-Centric Real-to-Sim-to-Real Learning from One Demonstration
Jun 29, 2026Heecheol Kim, Namiko Saito, Katsushi Ikeuchi +1Robot System Identification Contact-Rich Robotic Manipulation
Microsoft Research Asia - Tokyo · The University of Tokyo

Mar 6, 2026J/K move · Enter open · S save

Few-Shot Neural Differentiable Simulator: Real-to-Sim Rigid-Contact Modeling

Zhenhao Huang, Siyuan Luo, Bingyang Zhou, Ziqiu Zeng+2

National University of Singapore, Singapore · Prana Lab, San Francisco, USA

Differentiable Physics Robot Dynamics Learning Data-Efficient Robot Learning Synthetic Data Generation for Robotics Graph Neural Networks Contact-Rich Robotic Manipulation

Open paper AskPDF arXiv

Accurate physics simulation is essential for robotic learning and control, yet analytical simulators often fail to capture complex contact dynamics, while learning-based simulators typically require large amounts of costly real-world data. To bridge this gap, we propose a few-shot real-to-sim approach that combines the physical consistency of analytical formulations with the representational capacity of graph neural network (GNN)-based models. Using only a small amount of real-world data, our method calibrates analytical simulators to generate large-scale synthetic datasets that capture diverse contact interactions. On this foundation, we introduce a mesh-based GNN that implicitly models rigid-body forward dynamics and derive surrogate gradients for collision detection, achieving full differentiability. Experimental results demonstrate that our approach enables learning-based simulators to outperform differentiable baselines in replicating real-world trajectories. In addition, the differentiable design supports gradient-based optimization, which we validate through simulation-based policy learning in multi-object interaction scenarios. Extensive experiments show that our framework not only improves simulation fidelity with minimal supervision but also increases the efficiency of policy learning. Taken together, these findings suggest that differentiable simulation with few-shot real-world grounding provides a powerful direction for advancing future robotic manipulation and control.

Abstract

Sim-to-Real Transfer

Explore similar work

Few-Shot Neural Differentiable Simulator: Real-to-Sim Rigid-Contact Modeling

An Real-Sim-Real (RSR) Loop Framework for Generalizable Robotic Policy Transfer

ConCent: Contact-Centric Real-to-Sim-to-Real Learning from One Demonstration