Floating base robots gain safe real time MPC

A new three layer control stack lets floating base robots handle human contact at 1 kHz without steady state drift. The approach extends a fixed base impedance model predictive control framework into the floating realm. It stacks a centroidal MPC that plans contact forces over a 500 millisecond horizon, a priority driven whole body control layer that resolves balance into joint torques via a contact consistent null space projection, and a residual null space governed by a receding horizon quadratic program that predicts and rejects physical human robot interaction disturbances using a Kalman augmented state. A key move is a contact consistent feedback linearization that reduces the arm end effector to a double integrator with a constant state matrix within each contact mode, enabling offline precomputation of the QP cost and online operation at or above 1 kHz.

Testing shows this architecture can preserve stability and safety as a floating base shifts posture under external contact. By keeping a fixed effective model within contact modes, the system can switch between supports and still maintain a coherent impedance behavior. Documentation indicates a covariance inflation protocol preserves the disturbance estimate across contact mode switches, which is crucial for avoiding drift when the robot lurches between stance and swing or when a human partner applies a sustained force. In the infinite horizon, the team proves an impedance equivalence theorem: the three layer stack converges to a classical task space impedance law whose effective mass, damping and stiffness adapt to posture and contact configuration. The result is a control stack that behaves like a familiar impedance controller, but with a planner that anticipates contact forces and a feedback layer that remains balance aware under human contact.

The work is still at the lab and simulation stage, but the numbers are precise and the engineering decisions clear. Testing on a 17-DOF biped and on the Unitree G1 humanoid suggests the architecture scales beyond toy platforms. The paper reports simulations that demonstrate the envisioned online performance and stability properties across a set of contact scenarios, laying out a path from simulation to bench experiments and, eventually, cautious real world trials. For engineers, the most striking spec is the tight loop: a 1 kHz QP based residual controller paired with a 500 millisecond horizon planner, all while maintaining a contact consistent posture plan. For operators, the credibility rests on the disturbance handling; humans can push, pull, or lean on the robot, and the system is designed to keep the interaction compliant without letting the robot lose balance.

Two practical takeaways jump out for practitioners plotting a path to field use. First, there is a heavy emphasis on offline preparation: the QP cost can be precomputed, and the constant state matrix within each contact mode reduces online compute needs, but hardware choices will matter. The architecture benefits from fast solvers and robust state editors, or the latency budget will bite. Second, the approach hinges on reliable disturbance estimation and switch handling. The covariance inflation and Kalman augmentation are not decorative; they are essential to keep the interaction from producing steady state errors when contact configurations flip. That means real hardware will need careful sensor fusion and failure mode analysis to prevent misestimation from propagating into balance loss.

Looking ahead, observers will want real world demonstrations with multiple contact points and varied payloads, plus an explicit discussion of failure modes under noisy sensing or sudden, large pHRI impulses. The method's promise is clear: bring sophisticated impedance behavior to floating platforms without sacrificing real time responsiveness, even as humans touch the robot in dynamic, uncertain environments.