pith. sign in

arxiv: 2604.04409 · v2 · submitted 2026-04-06 · 💻 cs.RO · cs.MA

FORMULA: FORmation MPC with neUral barrier Learning for safety Assurance

Qintong Xie , Weishu Zhan , Peter Chin This is my paper

Pith reviewed 2026-05-10 20:17 UTC · model grok-4.3

classification 💻 cs.RO cs.MA
keywords multi-robot systemsformation controlmodel predictive controlcontrol barrier functionsneural networkssafety assurancedecentralized controlobstacle avoidance
0
0 comments X

The pith

FORMULA integrates model predictive control with trained neural barrier functions to let multi-robot teams hold formation while avoiding obstacles safely and at scale.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a control method for groups of robots that must keep a preset shape while moving through cluttered spaces filled with obstacles and other agents. Traditional predictive planning struggles to guarantee safety at large scales and requires experts to write complex safety rules by hand. FORMULA replaces the handcrafted rules with neural networks that learn to certify safety in a distributed fashion, then combines those networks with predictive planning and stability functions. If the method works, robot teams can navigate without collisions or formation loss and with lower onboard computation, opening the door to bigger deployments in logistics and emergency response.

Core claim

The authors introduce FORMULA as a distributed framework that pairs model predictive control for trajectory planning with control Lyapunov functions for stability and neural-network approximations of control barrier functions for safety. The learned barriers enforce decentralized collision avoidance and formation preservation without manual constraint design, resolve deadlocks in dense settings, and lower online computation relative to standard MPC while maintaining the desired geometric configuration.

What carries the argument

Neural network-based control barrier functions that are trained offline to supply decentralized safety certificates for nonlinear multi-robot dynamics.

If this is right

  • Large-scale multi-robot teams can navigate complex environments while preserving formation geometry.
  • Safety constraints no longer require expert handcrafting for each new scenario.
  • Deadlock situations in dense robot clusters are resolved through the learned barrier terms.
  • Online computational burden drops because the safety functions are evaluated by the trained network rather than solved as hard constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same learned-barrier idea could be tested on teams of aerial vehicles or ground vehicles with different dynamics.
  • Hardware experiments would likely expose sensitivity to sensor noise or communication latency not visible in simulation.
  • The training data for the neural barriers must cover edge cases such as sudden obstacle appearance or partial team failures.

Load-bearing premise

Neural networks can be trained to enforce safety constraints that reliably prevent collisions and formation breakup across the range of nonlinear robot motions and environments considered.

What would settle it

A simulation or hardware test in which robots controlled by FORMULA collide with each other or an obstacle, or lose the required formation spacing, would show the learned barriers fail to deliver the claimed safety.

Figures

Figures reproduced from arXiv: 2604.04409 by Peter Chin, Qintong Xie, Weishu Zhan.

Figure 1
Figure 1. Figure 1: Cooperative formation and safety-aware navigation for multiple mobile [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed FORMULA framework. Left: each robot runs a distributed MPC-CLF optimizer for formation stability and a NN–CBF [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Formation control in obstacle-dense environments. (a) Planar trajec [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Deadlock resolution in FORMULA. Colored trajectories show four [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

Multi-robot systems (MRS) are essential for large-scale applications such as disaster response, material transport, and warehouse logistics, yet ensuring robust, safety-aware formation control in cluttered and dynamic environments remains a major challenge. Existing model predictive control (MPC) approaches suffer from limitations in scalability and provable safety, while control barrier functions (CBFs), though principled for safety enforcement, are difficult to handcraft for large-scale nonlinear systems. This paper presents FORMULA, a safe distributed, learning-enhanced predictive control framework that integrates MPC with Control Lyapunov Functions (CLFs) for stability and neural network-based CBFs for decentralized safety, eliminating manual safety constraint design. This scheme maintains formation integrity during obstacle avoidance, resolves deadlocks in dense configurations, and reduces online computational load. Simulation results demonstrate that FORMULA enables scalable, safety-aware, formation-preserving navigation for multi-robot teams in complex environments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes FORMULA, a distributed MPC-based framework for multi-robot formation control that augments standard MPC with control Lyapunov functions (CLFs) for stability and neural-network approximations of control barrier functions (CBFs) for decentralized safety. The approach is claimed to eliminate manual CBF design, preserve formation geometry while avoiding obstacles, resolve deadlocks, and lower online computation, with simulation results asserted to show scalable, safety-aware navigation in cluttered environments.

Significance. If the learned neural CBFs can be shown to satisfy the invariance condition for the closed-loop nonlinear dynamics, the framework would offer a practical route to scalable safe formation control without handcrafted barriers, which is relevant for applications such as warehouse logistics and disaster response. The combination of MPC, CLFs, and learned CBFs is a natural direction, but its value hinges on whether the safety guarantees are actually delivered rather than merely asserted.

major comments (3)
  1. [Abstract and Section 4 (Neural CBF Training)] The central safety claim rests on neural CBFs enforcing invariance for the multi-robot dynamics, yet the manuscript supplies no post-training verification (analytic, SOS, or dense sampling) that the learned h(x) satisfies h(x) > 0 inside the safe set and L_f h + L_g h u + α(h) ≥ 0 for admissible states and controls. Without such verification, simulation success does not establish that the closed-loop trajectories remain inside the zero superlevel set, especially under formation-induced coupling or out-of-distribution states.
  2. [Abstract and Section 6 (Simulation Results)] The abstract asserts that FORMULA 'maintains formation integrity during obstacle avoidance' and 'resolves deadlocks in dense configurations,' but no quantitative metrics (e.g., formation error norms, deadlock resolution rates, or comparison against baseline MPC or handcrafted-CBF methods) are reported to support these claims. The simulation results therefore cannot be evaluated for whether they actually demonstrate the headline advantages.
  3. [Section 5 (MPC Formulation) and Section 6] The reduction in online computational load is listed as a benefit, but the manuscript does not report wall-clock times, solver iterations, or scaling curves with team size, nor does it compare against a non-learned CBF-MPC baseline. This leaves the scalability claim unsupported by concrete evidence.
minor comments (2)
  1. [Section 4] Notation for the neural CBF (e.g., how the network output is mapped to the barrier function h(x) and how the Lie derivatives are computed) should be defined explicitly with equations rather than left at the level of 'neural network-based CBFs.'
  2. [Section 4] The abstract mentions 'eliminating manual safety constraint design,' but the training loss or regularization terms used to encourage satisfaction of the CBF condition are not stated; adding these details would clarify the method.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will incorporate the suggested additions to strengthen the safety verification and empirical support.

read point-by-point responses

  1. Referee: [Abstract and Section 4 (Neural CBF Training)] The central safety claim rests on neural CBFs enforcing invariance for the multi-robot dynamics, yet the manuscript supplies no post-training verification (analytic, SOS, or dense sampling) that the learned h(x) satisfies h(x) > 0 inside the safe set and L_f h + L_g h u + α(h) ≥ 0 for admissible states and controls. Without such verification, simulation success does not establish that the closed-loop trajectories remain inside the zero superlevel set, especially under formation-induced coupling or out-of-distribution states.

    Authors: We acknowledge that the manuscript does not include explicit post-training verification of the CBF invariance conditions beyond the training loss. While the loss in Section 4 is constructed to encourage satisfaction of the CBF inequality, this does not provide a guarantee. In the revision we will add a verification subsection that performs dense sampling over the relevant state space (including coupled formation states and out-of-distribution points) and reports the fraction of states satisfying h(x) > 0 and the CBF derivative condition. We will also note the inherent limitations of learned approximations for formal guarantees. revision: yes

  2. Referee: [Abstract and Section 6 (Simulation Results)] The abstract asserts that FORMULA 'maintains formation integrity during obstacle avoidance' and 'resolves deadlocks in dense configurations,' but no quantitative metrics (e.g., formation error norms, deadlock resolution rates, or comparison against baseline MPC or handcrafted-CBF methods) are reported to support these claims. The simulation results therefore cannot be evaluated for whether they actually demonstrate the headline advantages.

    Authors: We agree that the current simulation section relies primarily on qualitative trajectory visualizations. To support the claims, the revised Section 6 will include quantitative tables reporting mean formation error norms, deadlock resolution success rates over repeated trials, and direct numerical comparisons against both standard MPC and handcrafted-CBF MPC baselines. revision: yes

  3. Referee: [Section 5 (MPC Formulation) and Section 6] The reduction in online computational load is listed as a benefit, but the manuscript does not report wall-clock times, solver iterations, or scaling curves with team size, nor does it compare against a non-learned CBF-MPC baseline. This leaves the scalability claim unsupported by concrete evidence.

    Authors: We recognize the need for concrete timing data. The revised manuscript will add wall-clock timing results, solver iteration counts, and scaling plots versus team size (e.g., 5–20 robots). We will also include a comparison against a non-learned CBF-MPC implementation to quantify the online computational savings. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; framework is an integration of established components

full rationale

The paper presents FORMULA as a distributed MPC framework augmented with CLFs for stability and neural-network CBFs for safety, with the explicit goal of eliminating manual barrier design. No equations in the abstract or described structure define a derived quantity in terms of itself, relabel fitted parameters as independent predictions, or rely on self-citations for uniqueness or load-bearing premises. The central claims rest on the integration and simulation validation rather than any reduction to tautological inputs. This matches the default case of a self-contained engineering contribution whose derivation chain does not collapse by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields limited visibility into parameters or assumptions; neural network training implies learned weights, and safety enforcement assumes learned barriers suffice for formal guarantees.

free parameters (1)
  • Neural network weights for CBF approximation
    Weights are learned to replace hand-crafted safety constraints, making them central to the decentralized safety claim.
axioms (1)
  • domain assumption Neural networks can be trained to produce valid control barrier functions that enforce safety in nonlinear multi-robot dynamics
    The elimination of manual design rests on this assumption for provable decentralized safety.

pith-pipeline@v0.9.0 · 5451 in / 1177 out tokens · 60840 ms · 2026-05-10T20:17:12.493138+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    Search and rescue under the forest canopy using multiple uavs,

    Y . Tian, K. Liu, K. Ok, L. Tran, D. Allen, N. Roy, and J. P. How, “Search and rescue under the forest canopy using multiple uavs,”The International Journal of Robotics Research, vol. 39, no. 10-11, pp. 1201– 1221, 2020

    work page 2020

  2. [2]

    Multiple uavs in forest fire fighting mission using particle swarm optimization,

    K. A. Ghamry, M. A. Kamel, and Y . Zhang, “Multiple uavs in forest fire fighting mission using particle swarm optimization,” in2017 Interna- tional conference on unmanned aircraft systems (ICUAS). IEEE, 2017, pp. 1404–1409

    work page 2017

  3. [3]

    Agent-based intelligent monitoring in large- scale continuous material transport,

    Y . Pang and G. Lodewijks, “Agent-based intelligent monitoring in large- scale continuous material transport,” inProceedings of 2012 9th IEEE International Conference on Networking, Sensing and Control. IEEE, 2012, pp. 79–84

    work page 2012

  4. [4]

    Double-deck multi-agent pickup and delivery: Multi- robot rearrangement in large-scale warehouses,

    B. Li and H. Ma, “Double-deck multi-agent pickup and delivery: Multi- robot rearrangement in large-scale warehouses,”IEEE Robotics and Automation Letters, vol. 8, no. 6, pp. 3701–3708, 2023

    work page 2023

  5. [5]

    Distributed optimization in multi-agent robotics for industry 4.0 warehouses,

    A. Kattepur, H. K. Rath, A. Simha, and A. Mukherjee, “Distributed optimization in multi-agent robotics for industry 4.0 warehouses,” in Proceedings of the 33rd Annual ACM Symposium on Applied Computing, 2018, pp. 808–815

    work page 2018

  6. [6]

    The before, during, and after of multi- robot deadlock,

    J. Grover, C. Liu, and K. Sycara, “The before, during, and after of multi- robot deadlock,”The International Journal of Robotics Research, vol. 42, no. 6, pp. 317–336, 2023

    work page 2023

  7. [7]

    Safety barrier certificates for collisions-free multirobot systems,

    L. Wang, A. D. Ames, and M. Egerstedt, “Safety barrier certificates for collisions-free multirobot systems,”IEEE Transactions on Robotics, vol. 33, no. 3, pp. 661–674, 2017

    work page 2017

  8. [8]

    Gcbf+: A neural graph control barrier function framework for distributed safe multi-agent control,

    S. Zhang, O. So, K. Garg, and C. Fan, “Gcbf+: A neural graph control barrier function framework for distributed safe multi-agent control,”IEEE Transactions on Robotics, 2025

    work page 2025

  9. [9]

    Formation control of leader–follower mobile robots’ systems using model predictive control based on neural- dynamic optimization,

    H. Xiao, Z. Li, and C. P. Chen, “Formation control of leader–follower mobile robots’ systems using model predictive control based on neural- dynamic optimization,”IEEE Transactions on Industrial Electronics, vol. 63, no. 9, pp. 5752–5762, 2016

    work page 2016

  10. [10]

    Agile coordination and assistive collision avoidance for quadrotor swarms using virtual structures,

    D. Zhou, Z. Wang, and M. Schwager, “Agile coordination and assistive collision avoidance for quadrotor swarms using virtual structures,”IEEE Transactions on Robotics, vol. 34, no. 4, pp. 916–923, 2018

    work page 2018

  11. [11]

    Information consensus in multivehicle cooperative control,

    W. Ren, R. W. Beard, and E. M. Atkins, “Information consensus in multivehicle cooperative control,”IEEE Control systems magazine, vol. 27, no. 2, pp. 71–82, 2007

    work page 2007

  12. [12]

    Formation tracking and obstacle avoidance for multiple quadrotors with static and dynamic obstacles,

    J. Qi, J. Guo, M. Wang, C. Wu, and Z. Ma, “Formation tracking and obstacle avoidance for multiple quadrotors with static and dynamic obstacles,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 1713–1720, 2022

    work page 2022

  13. [13]

    Usv formation and path-following control via deep reinforcement learning with random braking,

    Y . Zhao, Y . Ma, and S. Hu, “Usv formation and path-following control via deep reinforcement learning with random braking,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 12, pp. 5468–5478, 2021

    work page 2021

  14. [14]

    Nonlinear model predictive control of robotic systems with control lyapunov functions,

    R. Grandia, A. J. Taylor, A. Singletary, M. Hutter, and A. D. Ames, “Nonlinear model predictive control of robotic systems with control lyapunov functions,”arXiv preprint arXiv:2006.01229, 2020

    work page arXiv 2006

  15. [15]

    Adaptive clf-mpc with application to quadrupedal robots,

    M. V . Minniti, R. Grandia, F. Farshidian, and M. Hutter, “Adaptive clf-mpc with application to quadrupedal robots,”IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 565–572, 2021

    work page 2021

  16. [16]

    Distributed model predictive control for heterogeneous vehicle platoons under unidi- rectional topologies,

    Y . Zheng, S. E. Li, K. Li, F. Borrelli, and J. K. Hedrick, “Distributed model predictive control for heterogeneous vehicle platoons under unidi- rectional topologies,”IEEE Transactions on Control Systems Technology, vol. 25, no. 3, pp. 899–910, 2016

    work page 2016

  17. [17]

    Distributed nonlinear trajectory optimization for multi-robot motion planning,

    L. Ferranti, L. Lyons, R. R. Negenborn, T. Keviczky, and J. Alonso-Mora, “Distributed nonlinear trajectory optimization for multi-robot motion planning,”IEEE Transactions on Control Systems Technology, vol. 31, no. 2, pp. 809–824, 2022

    work page 2022

  18. [18]

    Distributed lyapunov-based model pre- dictive formation tracking control for autonomous underwater vehicles subject to disturbances,

    H. Wei, C. Shen, and Y . Shi, “Distributed lyapunov-based model pre- dictive formation tracking control for autonomous underwater vehicles subject to disturbances,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 51, no. 8, pp. 5198–5208, 2019

    work page 2019

  19. [19]

    Explicit distributed and localized model predictive control via system level synthesis,

    C. A. Alonso, N. Matni, and J. Anderson, “Explicit distributed and localized model predictive control via system level synthesis,” in2020 59th IEEE Conference on Decision and Control (CDC). IEEE, 2020, pp. 5606–5613

    work page 2020

  20. [20]

    Control barrier functions: Theory and applications,

    A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control barrier functions: Theory and applications,” in2019 18th European control conference (ECC). IEEE, 2019, pp. 3420–3431

    work page 2019

  21. [21]

    Correctness guarantees for the composition of lane keeping and adaptive cruise control,

    X. Xu, J. W. Grizzle, P. Tabuada, and A. D. Ames, “Correctness guarantees for the composition of lane keeping and adaptive cruise control,”IEEE Transactions on Automation Science and Engineering, vol. 15, no. 3, pp. 1216–1229, 2017

    work page 2017

  22. [22]

    Neural graph control barrier functions guided distributed collision-avoidance multi-agent control,

    S. Zhang, K. Garg, and C. Fan, “Neural graph control barrier functions guided distributed collision-avoidance multi-agent control,” inConference on robot learning. PMLR, 2023, pp. 2373–2392

    work page 2023

  23. [23]

    Learning safe multi-agent control with decentralized neural barrier certificates,

    Z. Qin, K. Zhang, Y . Chen, J. Chen, and C. Fan, “Learning safe multi- agent control with decentralized neural barrier certificates,”arXiv preprint arXiv:2101.05436, 2021

    work page arXiv 2021

  24. [24]

    Safe control with learned certificates: A survey of neural lyapunov, barrier, and contraction methods for robotics and control,

    C. Dawson, S. Gao, and C. Fan, “Safe control with learned certificates: A survey of neural lyapunov, barrier, and contraction methods for robotics and control,”IEEE Transactions on Robotics, vol. 39, no. 3, pp. 1749– 1767, 2023

    work page 2023

  25. [25]

    Safety-critical control and planning for obstacle avoidance between polytopes with control barrier functions,

    A. Thirugnanam, J. Zeng, and K. Sreenath, “Safety-critical control and planning for obstacle avoidance between polytopes with control barrier functions,” in2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 286–292

    work page 2022

  26. [26]

    Predictive control barrier func- tions: Enhanced safety mechanisms for learning-based control,

    K. P. Wabersich and M. N. Zeilinger, “Predictive control barrier func- tions: Enhanced safety mechanisms for learning-based control,”IEEE Transactions on Automatic Control, vol. 68, no. 5, pp. 2638–2651, 2022

    work page 2022

  27. [27]

    Guaranteed obstacle avoidance for multi-robot operations with limited actuation: A control barrier function approach,

    Y . Chen, A. Singletary, and A. D. Ames, “Guaranteed obstacle avoidance for multi-robot operations with limited actuation: A control barrier function approach,”IEEE Control Systems Letters, vol. 5, no. 1, pp. 127– 132, 2020

    work page 2020

  28. [28]

    Safe control synthesis via input constrained control barrier functions,

    D. R. Agrawal and D. Panagou, “Safe control synthesis via input constrained control barrier functions,” in2021 60th IEEE Conference on Decision and Control (CDC). IEEE, 2021, pp. 6113–6118

    work page 2021

  29. [29]

    Control barrier function based quadratic programs for safety critical systems,

    A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,”IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2016

    work page 2016

  30. [30]

    Clf-cbf based quadratic programs for safe motion control of nonholonomic mobile robots in presence of moving obstacles,

    M. Desai and A. Ghaffari, “Clf-cbf based quadratic programs for safe motion control of nonholonomic mobile robots in presence of moving obstacles,” in2022 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM). IEEE, 2022, pp. 16–21

    work page 2022