Stoch3

Quadrupedal robots possess the unique ability to navigate unstructured and hazardous terrains—from disaster zones to industrial inspection sites—where traditional wheeled robots often fail. However, achieving stable locomotion on hardware requires precise control strategies that can anticipate and react to dynamic physical constraints. The objective of this project was to deploy a Model Predictive Control (MPC) framework on the Stoch3 quadruped, enabling robust, dynamic walking in real-world conditions by optimizing control actions over a finite time horizon.

The core of the design involved translating a theoretical MPC formulation into a deployable software stack capable of running on the robot's onboard computer. While the controller performed well in simulation, the initial hardware integration revealed severe instability, preventing the robot from maintaining balance during trot gaits. I took the lead in the system-level debugging process, analyzing telemetry data to diagnose the discrepancy between the simulated and physical performance. My investigation isolated the root cause not within the controller itself, but deep within the state estimation pipeline, where sensor noise and latency were feeding inaccurate velocity estimates to the optimizer.

To resolve this, I re-engineered the state estimation filter to better fuse proprioceptive sensor data, significantly reducing drift and noise. Once the state estimator provided ground-truth quality data, the MPC controller immediately achieved stable locomotion, successfully tracking velocity commands and rejecting disturbances. This validation process was critical, as it confirmed that the control logic was sound and that the system's failure was a solvable integration issue rather than a fundamental theoretical flaw.

This project demonstrated the critical importance of accurate state estimation in high-performance robotics. By identifying and fixing the underlying estimation faults, I successfully bridged the "sim-to-real" gap, allowing the Stoch3 to utilize the full potential of MPC for stable navigation. My contribution ensured the robot's readiness for more complex tasks, laying the groundwork for future research into traversing uneven terrain and executing dynamic maneuvers.