Seminars

View all Seminars  |  Download ICal for this event

Learning Robust Legged Locomotion under Dynamic, Environmental, and Morphological Uncertainties 

Series: Ph.D. Colloquium

Speaker: Vamshi Kumar Kurva, Ph.D (Engg.) student, Dept. of CSA, IISc

Date/Time: Jul 20 11:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Faculty Advisor: Prof. Shishir N Y

Abstract:
Reliable deployment of legged robots in real-world environments requires robust locomotion under multiple sources of uncertainty. While locomotion with a fixed robot morphology on simple terrain is relatively well understood, practical deployment introduces changing payloads, uncertain inertial properties, challenging terrains, and variations in robot morphology. These uncertainties alter the underlying dynamics and make reliable velocity tracking increasingly difficult. This thesis investigates learning-based approaches for robust quadrupedal locomotion under progressively more challenging sources of uncertainty.

The first part considers locomotion under known dynamics on flat terrain. Instead of solving a large online optimization problem at every control step, a lightweight linear reinforcement learning policy directly generates high-level locomotion commands, substantially reducing computational complexity while maintaining robust velocity tracking. The second part addresses dynamic uncertainty caused by payload variations and centre of mass shifts through an adaptive model predictive control framework that estimates changing inertial properties online. The third part extends the problem to simultaneous dynamic and environmental uncertainties by developing an adaptive reinforcement learning framework that decomposes learning into successive stages, simplifying policy optimization while improving robustness to varying payloads and unstructured terrains, including slopes, stairs, and uneven surfaces. The final part addresses morphological uncertainty by learning a single visual world model across multiple robot morphologies, learning morphology-aware latent representations that support zero-shot transfer across robot configurations while maintaining robust locomotion over complex terrains.

Together, these works present a systematic study of quadrupedal locomotion under increasingly realistic deployment scenarios. Beginning with locomotion under known dynamics, the thesis progressively addresses dynamic, environmental, and morphological uncertainties encountered in practical applications. Collectively, these contributions demonstrate how learning can progressively improve the robustness and generality of quadrupedal locomotion under increasingly realistic sources of uncertainty. The proposed approaches are validated through extensive simulation and hardware experiments, demonstrating robust locomotion across diverse operating conditions.