Reproduce “Behavior Cloning of MPC for 3-DOF Robotic Manipulators” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Theo Guegan, Dexter Wen Jie Teo
The authors train a neural policy to replace inverse kinematics and model predictive control for a simulated three-degree-of-freedom arm reaching targets. It predicts torques from joint angles, velocities, and target position using expert demonstrations. Static multilayer perceptrons had lower test error than tested history-based models. MLP_Deep succeeded in 84.98% of episodes at five-centimeter tolerance, versus 64.16% at two-centimeter tolerance; results are from simulation.
Key insights
- Deep Multi-Layer Perceptrons (MLPs) consistently outperformed Recurrent Neural Networks (RNNs) and MLPs augmented with sliding windows, indicating that the immediate robot state provides sufficient information for the control decision.
- For MLP architectures, increasing network depth was found to be more beneficial than increasing width for accurately capturing the non-linear inverse dynamics of the robot.
- A precision gap under strict error tolerances (2.9 cm average final error) suggests that while behavior cloning captures the expert's global trajectory, it lacks MPC's continuous gradient-based error correction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.