Fast and Continuous Foothold Adaptation for Dynamic Locomotion through CNNs (1809.09759v2)

Abstract: Legged robots can outperform wheeled machines for most navigation tasks across unknown and rough terrains. For such tasks, visual feedback is a fundamental asset to provide robots with terrain-awareness. However, robust dynamic locomotion on difficult terrains with real-time performance guarantees remains a challenge. We present here a real-time, dynamic foothold adaptation strategy based on visual feedback. Our method adjusts the landing position of the feet in a fully reactive manner, using only on-board computers and sensors. The correction is computed and executed continuously along the swing phase trajectory of each leg. To efficiently adapt the landing position, we implement a self-supervised foothold classifier based on a Convolutional Neural Network (CNN). Our method results in an up to 200 times faster computation with respect to the full-blown heuristics. Our goal is to react to visual stimuli from the environment, bridging the gap between blind reactive locomotion and purely vision-based planning strategies. We assess the performance of our method on the dynamic quadruped robot HyQ, executing static and dynamic gaits (at speeds up to 0.5 m/s) in both simulated and real scenarios; the benefit of safe foothold adaptation is clearly demonstrated by the overall robot behavior.

• The paper introduces a CNN-based method to adjust robot foothold trajectories in real-time, achieving 200x computational efficiency improvement.
• It leverages self-supervised learning with over 17,000 training samples to achieve 99% step feasibility and a 76% optimal prediction rate.
• The approach enables quadrupedal robots to safely traverse rough terrains with agile mid-locomotion adjustments at speeds up to 0.5 m/s.

Analyzing CNN-Based Foothold Adaptation for Dynamic Locomotion

This paper by Villarreal et al. addresses the challenge of facilitating highly dynamic and adaptable legged robot locomotion across rough terrains using vision-based foothold adjustment strategies, complemented by onboard implementations of CNNs. The research presented focuses on leveraging convolutional neural networks to enhance the robustness and efficiency of real-time dynamic quadrupedal robot movement, particularly through the deployment of innovative foothold adaptation mechanisms.

The authors propose a CNN-based approach to dynamically adjust the trajectory of robot legs during the swing phase based on the perceived terrain, effectively providing a method to bridge the gap between blind reactive locomotion and vision-based planning strategies. The primary strength of this approach lies in its ability to operate autonomously and process complex terrain information rapidly, demonstrating computational speeds up to 200 times faster than prior heuristic-dependent methods. This efficiency is achieved through the use of self-supervised learning, where a heuristic algorithm generates ground truth data without reliance on human intervention, therefore enabling large-scale training datasets, which consist of 17,844 examples, in contrast to just 3,300 in previous works.

In terms of numerical performance, the CNN used for foothold adaptation reports 99% feasibility in selected steps, with an optimal prediction rate of 76%. The findings suggest that while perfect optimization in foothold placement isn't always achieved, the overall selection remains reliably safe according to key evaluative criteria. These criteria include kinematics, terrain roughness and uncertainty margins, frontal foot collisions, and leg collisions. This ensures substantial improvements in automated decision-making crucial for real-world deployment of quadruped robots.

The implications of this research are significant both theoretically and practically. From a theoretical perspective, this marks a meaningful advancement in deep learning applied within the field of robotics, demonstrating the transformative potential of CNNs to enhance real-time responses in complex environments. On a practical level, the methodology spells positive impacts on the deployment and control of robots designed for exploration and task execution in uncertain, rugged terrains, as evidenced by the experimental data showing successful gap-crossing initiatives at velocities up to 0.5 m/s. Notably, the system permits adaptable responses to external perturbations mid locomotion, mitigating the impacts through rapid adjustments in swing trajectory.

Looking forward, advancements such as real-time dynamic criteria incorporation, two-step horizon evaluations, and finer customization of CNN architectures could further refine predictive accuracy and locomotion efficiency. Such advancements would significantly augment robot autonomy and environmental interaction capacity, thus optimizing robotic applications in exploration, search and rescue missions, and terrestrial navigation within complex landscapes. Hence, this contribution not only reflects considerable progress in robotics locomotion strategies but also sets the stage for future explorations into AI-enhanced dynamic control systems in autonomous machines.