ReSCoT-Q

Reference-guided surface-frame control for quadrupeds — force-controlled tool sliding on Unitree Go1 hardware with MPC and whole-body control.

The surface-frame controller sliding the body-mounted tool along a cylindrical surface while regulating contact force (RViz visualization; commanded contact forces shown in magenta).

ReSCoT-Q (Reference-Guided Surface-Frame Control for Quadruped Robots) is my primary graduate research project in the ALMaS Research Group at WPI, advised by Dr. Mahdi Agheli. The goal: enable a quadruped to interact with the world through sustained contact — pressing a body-mounted tool against an external surface and sliding it along a reference path while regulating contact force, all while the robot keeps balancing on four legs.

This is a fundamentally different regime from locomotion or pick-and-place. The contact is persistent rather than intermittent, the interaction force must be actively regulated rather than avoided, and every stance transition of the gait perturbs the tool. Think of wiping a window, sanding a hull, or inspecting a pipeline — tasks that demand force-controlled loco-manipulation.

How it works

The controller reasons in a surface frame attached to the contact point, decomposing the task into three orthogonal objectives:

  • Normal direction — regulate the pressing force against the surface to a target (e.g., 10 N), with a normal-force lag state inside the MPC model so force dynamics are predicted, not just reacted to.
  • Tangential direction — track a commanded slide velocity along the reference path.
  • Binormal (cross-track) direction — stabilize lateral deviation from the path.

Three components make this robust on real hardware:

  1. Surface-frame MPC + weighted whole-body control. An OCS2 SQP centroidal MPC runs at 100 Hz, feeding a hierarchically weighted WBC that trades off contact objectives, gait execution, and balance on a Unitree Go1 with a body-mounted tool and a force/torque sensor.
  2. Confidence-gated surface estimation. A reference-guided estimator fuses a nominal surface prior (a rough CAD-like guess of the surface) with online contact-wrench and tool-motion measurements. Confidence gating decides when the online evidence should override the prior — the controller tolerates large geometric error in the assumed surface model.
  3. Contact-stability supervisor. A state machine gates sliding during contact acquisition, force recovery, and gait-induced contact dropouts, so transient force losses don’t destabilize the slide.
The same controller generalizes across surface geometry — here sliding against a planar wall.

Hardware results

On a cylindrical test surface, with the surface prior deliberately wrong by ~10% in radius (1.0 m prior vs. 0.903 m fitted cylinder):

Metric Result
Sustained contact duration 109.9 s
Surface coverage 91.3° arc of the cylinder
Normal-force RMS error 0.60 N at a 10 N target
Path-tracking RMS error 18.8 mm

Status

A first-author manuscript (S. Selvaraj and M. Agheli) is in preparation, targeting IEEE Robotics and Automation Letters (RA-L).