UC San Diego

Compared to wheeled robots, legged robots can provide a significant advantage in traversing complex, uneven terrain. The advantage is still unrealized because of the lack of effective algorithms that can give instantaneous reactions like biological systems. In recent years data-driven methods have gathered the community's interest due to the ability to learn rules/patterns from data. The black box data-driven methods remove the need for rule-based strategies while benefiting from the fast implementation. The data-driven method like Deep Reinforcement Learning (DRL) has the potential to make the system more reactive, efficient, and less reliant on the rules. The thesis examines the potential of DRL on rough terrain. While there is a lot of attention for reinforcement learning for legged robotics, the many aspects like environment generation, episode termination criteria are undiscussed in the literature available. We also suggest strategies of generating random terrain with appropriate parameters, environment design, and we also prove the use of randomization on the terrain to guarantee higher performance.

Due to imperfect estimation of terrain by the perception system, for legged locomotion, leg collisions with the terrain are inevitable. Due to the potential of damage, the collision is an important failure mode and needs to be investigated. The black box data-driven methods remove the need for rule-based strategies while benefiting from the fast implementation. In this thesis, we present a temporal convolutional network based CDLnet (Collision Detection and Localization Network), a single neural network both for collision detection and localization. Due to the unique nature of the problem, the conventional loss functions can raise a significant error in the measurements. Therefore a new loss function called CDLloss is introduced. The CDLloss uses a combination of classification loss and special collision localization loss. Due to the ability to effectively extract information across time-domain, the network outperforms the baseline neural network with the same loss function and also outperforms the best performing physics-based method on minitaur leg. Then we discuss future work directions and potential use-cases of this method.