Motion planning for off-road autonomous driving based on human-like cognition and weight adaptation

TL;DR

This paper tackles off-road motion planning by learning human-like environmental perception to adapt traversal costs. It introduces a multi-layer 2.5D feature map, a CNN-LSTM that predicts human-like trajectories, and a primitives-based planner that learns and updates cost weights $w=(w_h,w_r,w_\Gamma,w_s)$ to produce adaptive, human-like trajectories. Key contributions include hierarchical environment representation, a two-layer weight-learning objective using Mixture-of-Experts loss, and offline/online generation of behavioral primitives, validated in desert off-road experiments showing real-time, stable, and human-like planning across diverse scenarios. The approach promises improved robustness and applicability for autonomous off-road platforms by aligning planning with human-like cognition and vehicle dynamics.

Abstract

Driving in an off-road environment is challenging for autonomous vehicles due to the complex and varied terrain. To ensure stable and efficient travel, the vehicle requires consideration and balancing of environmental factors, such as undulations, roughness, and obstacles, to generate optimal trajectories that can adapt to changing scenarios. However, traditional motion planners often utilize a fixed cost function for trajectory optimization, making it difficult to adapt to different driving strategies in challenging irregular terrains and uncommon scenarios. To address these issues, we propose an adaptive motion planner based on human-like cognition and cost evaluation for off-road driving. First, we construct a multi-layer map describing different features of off-road terrains, including terrain elevation, roughness, obstacle, and artificial potential field map. Subsequently, we employ a CNN-LSTM network to learn the trajectories planned by human drivers in various off-road scenarios. Then, based on human-like generated trajectories in different environments, we design a primitive-based trajectory planner that aims to mimic human trajectories and cost weight selection, generating trajectories that are consistent with the dynamics of off-road vehicles. Finally, we compute optimal cost weights and select and extend behavioral primitives to generate highly adaptive, stable, and efficient trajectories. We validate the effectiveness of the proposed method through experiments in a desert off-road environment with complex terrain and varying road conditions. The experimental results show that the proposed human-like motion planner has excellent adaptability to different off-road conditions. It shows real-time operation, greater stability, and more human-like planning ability in diverse and challenging scenarios.

Figures (14)

• Figure 1: Flowchat. The left green box contains environmental point cloud data, global planning, and vehicle state. These inputs are fed into the human-like environment perception and learning framework located in the upper middle yellow box. This framework generates a traversal cost map and learns human-like trajectories using a CNN-LSTM network. The human-like results are then fed into the motion planning framework based on human-like cognition represented by the lower middle purple box, where the optimal cost weights are determined. Finally, these inputs are processed by the spatio-temporal coupling planning framework located in the right red box, resulting in the trajectory planning results with the optimal traversal cost weights.
• Figure 2: Three different resolutions of maps.
• Figure 3: The CNN-LSTM network learns about sequences of human driving trajectories. On the left, a sequence of environmental maps is input into the network sequentially over $T$ consecutive time steps. A CNN module is employed as the feature encoder, wrapping map information at each time step. The output is then fed into the LSTM module on the right, with the network's label representing the ground truth trajectory.
• Figure 4: Motion planning guided by human-like cognition. The green line represents the process of obtaining human-like trajectories based on environmental inputs. The blue line represents the process of generating clusters of primitive actions based on the ego's state. The yellow line represents the process of solving the cost weights associated with human-like behaviors and ultimately generating the spatiotemporal motion planning results.
• Figure 5: The vehicle platform and various testing scenarios.