Stoch3

Meet the hardware
( : Own Hardware = Best Hardware : )

Introducing the Stoch3/4/5 Lineup.
A custom fully in-house built quadruped platform.

It's an amazing experience working with custom hardware.
It's painfull most of times, but equally gratifying sometimes.
You are free to break it, with the only consquence that you'll yourself be the one fixing it later...
It's a non-stop cycle of find issue, report issue, work around issue, fix issue.

The Locmotion Control Stack

State Estimation:
IMU + proprioception + contaxt_estimation + slope estimation

Swing Controller:
Raibert Swing Controller

Stance Controller:
ConvexMPC

ConvexMPC

During stance phase, we calculate for the desired ground reaction forces (GRF's).

Objective: L2 loss wrt reference velocity trajectory + regularizers
Dynamics Constrains: Linearized single rigid-body-model (SRBM)
Friction Constrains: Convex cone-based no-slip conditions.

To the right are my notes on the linearized SRBM model and QP formulation of MPC we used.

MPC_cropped_removed.pdf

Testing Robustness

Different payload and COM distributions by attaching floating mass to the urdf

Testing on programatically developed clean slopes and obstacles

Testing and tuning to reject external forces and disturbances

Real2Sim?

To develop and environemnt to verify if our controller worked on the desired mountainous regions,
we headed out to out mountains,
with a small box with LIDAR and a depth camera,
recorded the data stream for our entire hike,
processed filtered the resulting point clouds,
and then converted and loaded the mesh to our favourite simulators for testing.