Augmenting Differentiable Simulators with Neural Networks to Close the Sim2Real Gap

TL;DR

Problem: bridging the sim2real gap in articulated rigid-body dynamics. Approach: a differentiable, hybrid simulator that can augment analytical models with data-driven residuals and supports end-to-end backpropagation of gradients to both analytical parameters and neural weights, via $\theta = [\theta_{AM}, \theta_{NN}]$. Key contributions: a neural-augmented simulation framework with neural-blueprint-based augmentation, a system-identification objective $\mathcal{L} = \sum_t ||f_\theta(s_{t-1}) - s^*_t||^2 + R||\theta_{NN}||^2$, and a parallel basin hopping strategy to escape poor local minima. Findings: preliminary sim2sim and real-data experiments reduce the gap between simulated and target trajectories, demonstrating practical viability for planning and control in robotics. Significance: by enabling gradient-based optimization over both analytic dynamics and learned residuals, this work offers a scalable path to more accurate simulators, with future directions including physics-constrained network architectures.

Abstract

We present a differentiable simulation architecture for articulated rigid-body dynamics that enables the augmentation of analytical models with neural networks at any point of the computation. Through gradient-based optimization, identification of the simulation parameters and network weights is performed efficiently in preliminary experiments on a real-world dataset and in sim2sim transfer applications, while poor local optima are overcome through a random search approach.

Figures (2)

• Figure 1: Left: Trajectories from rigid body contact simulation of a cube thrown to the right. Starting with poor model parameters, the box falls through the ground (blue). After optimizing \ref{['eq:sim2rel_loss']}, our simulation (orange) closely matches the target trajectory (green). Right: After system identification of a real double pendulum asseman2018learning, the sim2real gap is strongly reduced.
• Figure 2: Left: comparison of various model architectures (cf. Anurag et al. anurag2018hybrid). Right: augmentation of differentiable simulators with our proposed neural scalar type where variable $e$ becomes a combination of an analytical model $\phi(\cdot,\cdot)$ with inputs $a$ and $b$, and a neural network whose inputs are $a$, $c$, and $d$.