Learning deformable linear object dynamics from a single trajectory

Dataset

Code

The manipulation of deformable linear objects (DLOs) via model-based control requires an accurate and computationally efficient dynamics model. Yet, data-driven DLO dynamics models require large training data sets while their predictions often do not generalize, whereas physics-based models rely on good approximations of physical phenomena and often lack accuracy. To address these challenges, we propose a physics-informed neural ODE capable of predicting agile movements with significantly less data and hyper-parameter tuning. In particular, we model DLOs as serial chains of rigid bodies interconnected by passive elastic joints in which interaction forces are predicted by neural networks. The proposed model accurately predicts the motion of an robotically-actuated aluminium rod and an elastic foam cylinder after being trained on only thirty seconds of data.

Method

We approximate a DLO as a serial chain of pseudo-rigid bodies connected by two degrees of freedom in elastic joints. As a result, the positions and velocities of the elastic joints become the latent state x. Furthermore, we obtain a decoder in the form of forward kinematics that maps the latent state x and inputs u — the motion of the DLO's controlled end — into outputs (observations) y: y=h(x,u).

This approximation allows us to derive the dynamics of the DLO via the dynamics of the pseudo-rigid body chain combined with torques generated by elastic joints. These torques are classically modeled as a linear spring and damper. To obtain higher accuracy and model the material nonlinearities, we propose using neural networks for modeling the torques at elastic joints and call it neural pseudo-rigid body approximation (NPRBA). The figure on the left schematically describes the proposed method.

Results

When trained on a single trajectory, the proposed model outperforms the physics-based vanilla pseudo-rigid body method (VPRBA) and data-driven models: linear time invariant dynamics model (LTI), second-order neural ODE (NODE), and PRB-Net.

All the renderings of the test rollouts are at x0.2 speed. Note that the shape is reconstructed from latent state that in turn was estimated from the partial observations.

Page updated

Google Sites

Report abuse