Speeding Up Optimization-based Motion Planning through Deep Learning

TL;DR

This work tackles fast, collision-free motion planning for high-DoF robots in complex 3D environments by learning to predict near-optimal trajectory intermediates that warm-start an optimization-based planner. It introduces Basis Point Set (BPS) as an efficient, permutation-invariant world encoding and couples it with a data-driven training regime that balances challenging and consistent samples through cleaning, boosting, and extending of the dataset. A CHOMP-like planner with explicit swept-volume collision checking and a gradient-descent solver forms the optimization backbone, while the network provides a strong initial guess that drastically accelerates convergence. The approach enables rapid planning for the 19-DoF Agile Justin (around 200 ms on a single CPU) and demonstrates a first successful sim-to-real transfer, highlighting practical impact for real-time reactive planning in unseen environments.

Abstract

Planning collision-free motions for robots with many degrees of freedom is challenging in environments with complex obstacle geometries. Recent work introduced the idea of speeding up the planning by encoding prior experience of successful motion plans in a neural network. However, this "neural motion planning" did not scale to complex robots in unseen 3D environments as needed for real-world applications. Here, we introduce "basis point set", well-known in computer vision, to neural motion planning as a modern compact environment encoding enabling efficient supervised training networks that generalize well over diverse 3D worlds. Combined with a new elaborate training scheme, we reach a planning success rate of 100%. We use the network to predict an educated initial guess for an optimization-based planner (OMP), which quickly converges to a feasible solution, massively outperforming random multi-starts when tested on previously unseen environments. For the DLR humanoid Agile Justin with 19DoF and in challenging obstacle environments, optimal paths can be generated in 200ms using only a single CPU core. We also show a first successful real-world experiment based on a high-resolution world model from an integrated 3D sensor.

Figures (10)

• Figure 1: DLR's Agile Justin Bauml2014 in a challenging obstacle environment, we generate with simplex noise for training and testing. The sphere model of the 19 DoF humanoid is shown as red wireframes. For videos of the motions, visit \gitlink.
• Figure 2: Three different robots in random 2D and 3D environments, generated with simplex noise. A motion problem is described by the world (hatched), the start configuration (blue), and the goal configuration (red). The solution, the shortest feasible path from start to goal, is drawn in gray. See \ref{['tab:datasets']} for an overview of the different robots and worlds in numbers.
• Figure 3: Substeps $q_{t, u}$ between two discrete waypoints $q_{t}$ and $q_{t+1}$ to explicitly calculate the swept volume of the path in higher resolution.
• Figure 4: Network architecture to map from a given motion task (world, start and end configuration) to an optimal path. Blocks of tapered Fully Connected Layers (gray) are combined like the DenseNet architecture Jegou2016 via skip-connections and concatenations. See the bottom of \ref{['tab:datasets']} for the number of network parameters used for the different robots.
• Figure 5: Scheme showing the connection between non-learning-based solver, dataset, and neural network. The colored arrows indicate the information flow for cleaning, extending, and boosting the dataset with the guidance of the network. Cleaning: use the solver to update labels in the solutions set; Boost: overrepresent existing hard examples in training; Extend: generate new hard examples for the network and add them to the problem set.