Kurate.org
Rankings
Back to Rankings

Safe-Ballbot: Tilt-Consistent MPC with Control Barrier Functions for Ballbot Navigation

Duc Cuong Vu, Tung Lam Nguyen

Nov 2, 2026arXiv:2611.03307v1
cs.ROeess.SY
Share
Scorecard· 16/16
5.5/10 impact

Solid, well-validated integration with a neat planner-tracker consistency insight, but incremental over the authors' prior work, simulation-only, and confined to a niche platform.

Abstract

Ball-balancing robots (ballbots) must lean to accelerate, and the lean is limited by the wheel torques and by safety. Planners that use the full robot model are expensive, while planners that ignore the balance dynamics produce references that cannot be followed. This paper presents a two-layer navigation scheme for a three-omni-wheel ballbot. The upper layer is a model predictive controller on a \emph{jerk model} of the ball position, with discrete-time control barrier functions for static, moving, rotating and concave obstacles. The lower layer is a hierarchical sliding mode controller with an extended state observer and a robust exponential control barrier function on the tilt (HSMC-ESO-ECBF). The layers are linked by an acceleration bound derived in closed form from the largest steady tilt that the barrier function admits, so that the plan never asks for a lean the tracker refuses. For obstacles whose motion is only predicted, the obstacles are inflated by conformal prediction radii that bound the prediction error with a prescribed probability. In MuJoCo, with omni-wheel roller contacts and torque saturation, the scheme keeps at least 1010\,cm of clearance, tracks the plan within 22\,cm RMS, keeps the tilt below the limit and plans in 99--1212\,ms, 1010--1515 times faster than a full-model nonlinear MPC that enters the inflated obstacles; with predicted obstacles, the conformal inflation removes the collisions of a constant-velocity predictor in 3030 random crossing scenarios and reduces them in an endless run on a closed path.The source code and video illustrations are available at https://safe-ballbot.github.io

AI Impact Assessments

(1 model)

Core Contribution

The paper proposes a two-layer navigation stack for a three-omni-wheel ballbot. The upper layer is a jerk-model (triple-integrator) MPC with discrete-time CBFs for static, moving, rotating and concave (ellipse-decomposed) obstacles. The lower layer is an HSMC-ESO-ECBF tracker that enforces a tilt cone. The central idea is the planner–tracker interface. The robust ECBF (with a 2δ‖φ‖₁ term) does not admit every lean up to θ_max. Proposition 2 gives the largest steady admitted tilt φ⋆ in closed form, and ¯φ⋆ for the worst direction. The planner's acceleration bound is derived from φ⋆ rather than from θ_max, so it never requests a lean the tracker will refuse. A secondary contribution adds conformal-prediction radii to inflate obstacles whose motion is only predicted, with soft barriers and a fallback plan for feasibility. Other pieces are a closed-form pole-placement sliding surface (Prop. 1) and a cross-validation of planar, 3D (Kane) and MuJoCo models.

The idea is neat and well motivated, but it is mostly an integration of known components: jerk MPC, DCBF, ECBF, ESO, split conformal prediction (following Lindemann et al.), and ellipse decomposition (following Pham et al.). The new element is the observation and closed-form quantification that the robust barrier shrinks the steady-state feasible lean set. Its analysis is elementary, a quadratic root, though it is correct and useful.

Methodological Rigor

  • The derivations are correct and clearly presented. The Prop. 2 proof is sound, and the DCBF induction argument is standard.
  • The simulation is relatively demanding for this subfield. It uses a MuJoCo model with roller contacts and torque saturation that differs from the design model, and it measures a 30% input-gain mismatch.
  • Prop. 2 is checked empirically (Fig. 9). The held tilt saturates near the predicted values but exceeds them by 0.1–0.3°, an honest discrepancy that the authors attribute to the gain mismatch.
  • The ablations are informative: the θ_max-derived bound raises RMS tracking error 1.8–7× and causes one goal failure, and removing the tracking margin collapses clearance.
  • The full-model NMPC comparison is partly favorable to the proposed scheme. The NMPC plans without the tilt limit and is filtered afterwards, which is a deliberately weak configuration, although the paper adds variants with no filter and a longer horizon. The NMPC's reported solve times of 108–150 ms exceed its 20 ms period, which makes it an infeasible baseline for real-time use.
  • There are 50 Monte Carlo runs with 11 clearance violations of the tracking margin, and a model-mismatch study.
  • The conformal evaluation is modest: 30 scenarios, single runs in Scenarios 3 and 4, and no statistical testing. The coverage guarantee is per step and per obstacle, not per run, and exchangeability is only approximately satisfied (correlated scores). The conformal inflation does not remove collisions in the endless run (3 → 1), and it fails on sharply out-of-distribution maneuvers. The authors acknowledge these limits candidly.
  • The r_track margin is empirical rather than a certified tracking-error bound, so there is no formal end-to-end safety guarantee. The planner starts from the previous plan rather than the measured state.
  • Everything is simulation only, with no hardware.

    • Potential Impact

  • The direct audience is small: ballbot control and planning researchers, and by extension groups working on dynamically balancing or underactuated platforms.
  • The conceptual lesson generalizes. When a tracker's safety filter restricts the reachable state set more than the nominal constraint does, the planner should use the tracker-admitted set. The same issue arises in quadrotor attitude limits, legged and humanoid robots, and any layered CBF-based stack.
  • The jerk-model planner with a 10 ms solve is practical and reproducible, and the released code and MuJoCo model add value.
  • Overall impact is likely incremental to moderate: useful within the niche, but unlikely to reshape wider robotics.

    • Timeliness & Relevance

    Layered safe planning and control, CBF-MPC, and conformal prediction for planning are all active topics. The ballbot niche is small, but interest has grown (asRoBallet, RL for ballbots). Combining CBFs with conformal prediction is fashionable but not new.

    Strengths

  • A clear, closed-form planner–tracker consistency insight that is quantitatively validated by ablation.
  • A realistic simulation with unmodeled dynamics, plus model-mismatch and Monte Carlo robustness checks.
  • Honest reporting of failures and limitations.
  • Roughly 10–15× computational savings over the full-model NMPC.
  • Good writing and a table of related work.
  • Code is released.

    • Limitations

  • Strongly incremental on the authors' own prior work ([5], [6]), with many components reused.
  • No hardware validation, so the sim-to-real gap is untested.
  • No formal end-to-end safety guarantee, and the empirical r_track is hand-tuned.
  • Conformal guarantees are weakened in practice: per-step only, with approximate exchangeability, and collisions persist in the long run.
  • A small number of test scenarios for the predicted-obstacle part, with single-run results.
  • The baseline NMPC is configured in a way that favors the proposed method.
  • The result depends on one specific robot and tracker design (δ, ℓ₁ choices).

    • Other Observations

    The paper is dense, covering many contributions at once (tracker, planner, interface, conformal extension, model validation), which dilutes the headline contribution. Reproducibility is good because parameters are fully listed and the model is public. The contribution is mostly an engineering and control-theory integration rather than a conceptual breakthrough.

    Rating:5.5/ 10
    Significance 4.5Rigor 6.5Novelty 5Clarity 7.5

    Generated Nov 3, 2026

    Comparison History (0)

    No comparisons yet.