CableRobotGraphSim: A Graph Neural Network for Modeling Partially Observable Cable-Driven Robot Dynamics

TL;DR

A novel Graph Neural Network model for cable-driven robots that aims to address shortcomings of prior simulation solutions by representing cable-driven robots as graphs, with the rigid-bodies as nodes and the cables and contacts as edges, which can quickly and accurately match the properties of other simulation models and real robots.

Abstract

General-purpose simulators have accelerated the development of robots. Traditional simulators based on first-principles, however, typically require full-state observability or depend on parameter search for system identification. This work presents \texttt{CableRobotGraphSim}, a novel Graph Neural Network (GNN) model for cable-driven robots that aims to address shortcomings of prior simulation solutions. By representing cable-driven robots as graphs, with the rigid-bodies as nodes and the cables and contacts as edges, this model can quickly and accurately match the properties of other simulation models and real robots, while ingesting only partially observable inputs. Accompanying the GNN model is a sim-and-real co-training procedure that promotes generalization and robustness to noisy real data. This model is further integrated with a Model Predictive Path Integral (MPPI) controller for closed-loop navigation, which showcases the model's speed and accuracy.

Figures (9)

• Figure 1: Left: Static, open-source 3-bar tensegrity platform. Right: The platform rolling clockwise.
• Figure 2: The three stages of a simulation step in the proposed learning process. Left: Graph and feature generation, where the robot's state $\mathbf{X}_t$, rest length $L_t^{rest}$, a history of controls $U_{t-h:t}$ and the future controls $U_{t+1:t+n}$ are used to construct the current graph $\mathcal{G}_t$ made up of nodes $\mathcal{N}_t$, edges $\mathcal{E}_t$ and their feature vectors. Middle:$\mathcal{G}_t$ is passed to MLP encoders and then to an LSTM recurrent block along with a hidden state $H_t$ and memory $C_t$ to output the updated hidden state $H_{t+1}$ and memory $C_{t+1}$. $H_{t+1}$ is passed to the processor, where it goes through $L$ rounds of message passing. The latest latent node and cable edges vectors are passed to the decoders to predict $n$ changes in velocities and rest lengths. Right: Multi-step time integration, where the decoders' output are used to iteratively integrate the rest lengths $L_t^{rest}$ to $L_{t+n}^{rest}$ and state $\mathbf{X}_t$ to $\mathbf{X}_{t+n}$.
• Figure 3: Sim-and-real co-training. Multiple simulation datasets are generated by varying system parameters, and are pooled with a real dataset of unknown system parameters. Data points are distinguished by a one-hot encoding node feature of the dataset enumerated index.
• Figure 4: Sim2sim evaluation for full trajectories and short-horizons, as well as a 3-bar and 6-bar tensegrity.
• Figure 5: Running MPPI controller in a loop of model training, task execution, and new data generation for model retraining over 3 iterations. Positional and rotational errors are evaluated on a pooled test set from all iterations. Success rate and task completion time are measured per iteration.