Imagine you're carrying a full cup of coffee across a crowded room. You can't plan your path by thinking about the fluid dynamics of the coffee — that's too slow. Instead, you internalize one rule: "don't tilt more than this much." That single constraint lets you dodge people quickly without spilling. This paper does exactly that for a ball-balancing robot: it derives a closed-form bound on how much the robot can lean, and hands that bound to a fast trajectory planner that never asks the balance controller to do the impossible. The committed claim: a two-layer navigation architecture for a three-omni-wheel ballbot that keeps trajectory planning in the 9–12 ms range while guaranteeing safety constraints — obstacle clearance ≥10 cm, tracking error ≤2 cm RMS, and tilt always below the physical limit — by linking layers through a provable acceleration bound derived from control barrier functions (CBFs). This is not the first use of CBFs in mobile robotics, but the specific contribution is the closed-form coupling between the planning and tracking layers that makes the two-layer decomposition formally consistent rather than heuristic. The architecture stacks two well-understood families. The upper layer runs model predictive control (MPC) on a jerk model of ball position — a triple integrator, cheap to solve. Discrete-time control barrier functions handle static, moving, rotating, and concave obstacles within the MPC. The lower layer is a hierarchical sliding mode controller (HSMC) with an extended state observer (ESO) and a robust exponential CBF that enforces the tilt constraint. The key connector is analytical: the authors derive the maximum steady-state tilt the ECBF will allow, convert that to a maximum ball acceleration, and impose that as a constraint in the upper-layer MPC. The plan can never demand a lean the tracker will refuse. For moving obstacles whose future positions are predicted, the authors use conformal prediction to inflate obstacle radii such that the prediction error is bounded with a user-specified probability. This is a clean statistical tool: instead of hand-tuning safety margins, you get margins calibrated from data with coverage guarantees. In 30 random crossing scenarios in MuJoCo with omni-wheel roller contacts and torque saturation, the conformal inflation eliminated all collisions that a naive constant-velocity predictor produced, and reduced collisions in an endless closed-path run. The comparison baseline is a full-model nonlinear MPC that plans directly over the ballbot's coupled dynamics. That planner is 10–15× slower (planning times not in the 9–12 ms range) and enters inflated obstacle regions. The proposed scheme avoids both problems. However, the entire validation is in MuJoCo simulation — no hardware experiments. The physics are reasonably rich (roller contacts, torque saturation), but sim-to-real transfer for a ballbot with three omni-wheels is non-trivial. The authors acknowledge this implicitly by releasing code and videos but not hardware results. The field fight here is between full-model planners — which are correct but slow — and decomposed planners that are fast but typically lack formal guarantees that the layers are consistent. This paper lands firmly on Side B: decompose, but close the gap with a provable coupling constraint. The next milestone is straightforward: demonstrate this on physical hardware with real sensor noise, actuator delay, and floor irregularities. The 9–12 ms planning cycle is well within real-time requirements, so the compute story transfers; the open question is whether the ESO and ECBF handle unmodeled dynamics robustly enough. The authors almost certainly plan hardware validation next — this reads like a theory-and-sim paper establishing the formal framework before a hardware companion paper.
