Source-linked AI summary
Distributed Predictive Flocking for a Swarm of Miniature Vibration-Driven Robots in a Confined Arena
Prajol Shrestha, Siddhant Mohapatra, Filip Novkoski, Julian Oelhaf, Nicolas Vandewalle, Ana-Sunčana Smith, Andreas Maier
TL;DR
Dense miniature-robot swarms in confined arenas face a collision-avoidance challenge, and the effective safety formulation may determine how many robots can coexist. The paper compares six formulations in a distributed predictive flocking controller and finds the Look-ahead Slack Dt-CBF most effective at preventing collisions, with high inter-robot margins across most tested swarm sizes.
Problem
In confined, high-density swarms, it remains open which inter-robot safety formulation lets the most robots coexist with minimal collisions.
Source: §1 · p. 3Method
Each robot solves a local NMPC combining flocking, task objectives, and safety constraints, exchanging predicted trajectories with neighbors; six safety formulations are compared under identical conditions.
Source: §1 · p. 3, §2.5.3 · p. 8Results
The LA Slack Dt-CBF is most effective at preventing collisions and provides the highest inter-robot margins for most swarm sizes tested.
Source: §4 · p. 20Takeaways & Limitations
The same controller architecture supports phototaxis, phototactic orbiting, and flocking without a stimulus through local optimization of cohesion and alignment costs coupled with safety constraints.
Source: §4 · p. 20The current single-process simulation advances robots sequentially, creating a wall-time bottleneck that the authors leave for future parallel implementation.
Source: §3.6 · p. 20
Abstract
from arXiv · showhide
Coordinating many miniature robots in a confined arena, with no leader, central planner, or global map, is a core challenge for swarm robotics. As the swarm grows denser, collision avoidance becomes the limiting factor, and the number of safely coexisting robots depends on the formulation of the inter-robot safety constraint. We present a distributed predictive flocking controller in which each robot solves its own nonlinear model predictive control (NMPC) problem, optimizing swarm cohesion and alignment costs, inter-robot and obstacle safety, and a task objective, while exchanging only predicted trajectories with local neighbors. Within this generalized controller architecture, solved in real time, we compare six barrier and control-barrier-function (CBF) safety formulations on a digital twin of an open-source vibration-driven bristlebot in a $0.64\,\mathrm{m^2}$ arena. A look-ahead discrete-time CBF performs best overall, with the fewest collisions and the widest inter-robot margins across nearly every swarm size tested. We deploy it on phototaxis, phototactic orbiting, and cohesion without a stimulus tasks, using swarm sizes starting from ten robots and up to the geometric packing limit. The swarm holds well: collisions stay low in small swarms, rising steeply as the packing limit approaches, while coordination and cohesion degrade gently rather than collapsing. The swarm is found not to trade safety for order, keeping both across the confined arena, with local alignment high at every swarm size.
1. Introduction
The paper targets dense, confined swarms where collision avoidance constrains how many robots can coexist, and asks which safety formulation best supports flocking. It introduces a distributed predictive controller and compares six safety formulations while characterizing swarm behavior near the arena’s packing limit.
1. Introduction
- In confined, cluttered arenas, increasing swarm density hinders smooth motion and makes collision avoidance a central challenge.Miniature robots also face tight onboard compute, memory, and power constraints.Source: §1 · p. 2
- The open question is which safety formulation lets the most robots coexist with minimal collisions in a fixed, confined, cluttered arena.Prior flocking NMPC work largely assumes sparse swarms in large airspace, while each work fixes one safety formulation.Source: §1 · p. 3
- The study compares six inter-robot safety formulations within one controller, varying only the safety constraint.The slacked barrier paired with a discrete-time CBF at a forward handle sustains the highest density with the fewest collisions.Source: §1 · p. 3
- The controller gives each robot a local NMPC combining Reynolds cohesion, alignment, separation, inter-robot and obstacle safety, and a task objective.Robots exchange predicted trajectories with selected neighbors, without a leader, central planner, or global map.Source: §1 · p. 3
- The controller generalizes across phototaxis, phototactic orbiting, and flocking without a stimulus, without modifying its safety or coordination layers.The paper frames task objectives and coordination structure as separable.Source: §1 · p. 3
- The Look-ahead Slack Dt-CBF gives the best overall trade-off for phototaxis, while collisions rise steeply near the geometric packing limit and coordination degrades gently.The controller retains cohesion with few collisions in small swarms; coordination and feasibility degrade rather than collapse as saturation sets in.Source: §1 · pp. 3–4
2. Preliminaries and Problem Statement
The problem is framed as a swarm of robots in a bounded arena, with local communication, nearby obstacles, and a light stimulus. The model uses elliptical robot footprints, unicycle dynamics, and a Gaussian light field that may be stationary or orbiting.
2. Preliminaries and Problem Statement
- Each robot selects up to K nearest neighbors within communication radius R_comm and accounts for nearby obstacles and a local light source.The directed sensing graph defines which robots can observe one another.Source: §2 · p. 4
2.1. Robot Model
- Each robot is modeled as a rigid ellipse matching the physical GRASPion bristlebot footprint, and its motion follows a three-state kinematic unicycle model.The continuous-time dynamics are integrated with ERK-4 using three steps per control interval.Source: §2.1 · p. 4
2.2. Light Field Model
- The light field is a Gaussian intensity profile whose source is either stationary at the arena center or orbits it at fixed radius and angular speed.Its analytic intensity gradient provides the local cue used to form behavior-mode references.Source: §2.2 · pp. 4–5
2.3. Control Architecture Overview
Each robot senses locally, generates a behavior-dependent look-ahead reference, and solves its own NMPC problem. It exchanges predicted trajectories with neighbors and synchronously replans without a central node.
2.3. Control Architecture Overview
- Flocking costs promote cohesion and heading alignment, while barrier and control-barrier constraints enforce inter-robot and obstacle safety.Robots publish predicted trajectories and re-solve synchronously using their neighbors’ previous plans.Source: §2.3 · p. 5
- Perception estimates each robot’s position, orientation, and velocity; planning uses that state to generate a behavior-specific reference and compute NMPC commands.A calibration layer passes commands to the actuator, while communication filters incoming broadcasts by range and nearest-neighbor selection.Source: Figure 2 · p. 6
2.4. Collective Behavior Modes
A behavior module converts local sensory observations into a look-ahead reference that the OCP tracks while maintaining flocking coherence and safety. The paper describes phototaxis and orbital behavior, and explains how other tasks can be incorporated through the same interface.
2.4. Collective Behavior Modes
- Collective behavior modes map local light-field observations into look-ahead position, heading, velocity, and angular-velocity references for the OCP.The reference position is clipped to remain inside the arena.Source: §2.4 · p. 6
2.4.1. Phototaxis Behavior Mode
- Phototaxis aligns each robot’s reference heading with the light-intensity gradient’s steepest ascent, without requiring knowledge of the source’s global position.When the gradient is nearly flat, the robot holds its current heading rather than turning randomly.Source: §2.4.1 · pp. 6–7
2.4.2. Phototactic Orbital Behavior Mode
- In phototactic orbital behavior, robots follow the tangent to an intensity level set, with a circulation parameter selecting clockwise or counter-clockwise motion.The reference heading is obtained by rotating the gradient direction by π/2.Source: §2.4.2 · p. 7
2.4.3. Extensibility to Arbitrary Behaviors
- The reference generator can support other behaviors by mapping local observations to look-ahead references while the DMPC retains safety and flocking coherence.Examples listed include directed migration, dispersal, transport, and shape formation.Source: §2.4.3 · p. 7
2.5. Reynolds Flocking Rules
The controller implements local Reynolds flocking by combining cohesion and alignment costs with barrier-based separation and obstacle constraints. Six inter-robot safety formulations vary the barrier and rate condition, while obstacle avoidance reuses the selected formulation.
2.5.1. Cohesion; 2.5.2. Alignment
- Cohesion and alignment are quadratic per-robot costs, enabled only when a neighbor is present, while separation and obstacle avoidance use barrier or discrete-time CBF constraints.Cohesion uses stage-matched predicted neighbor positions; alignment uses predicted neighbor velocities, with faster neighbors weighted more heavily.Source: §2.5 · p. 7, §2.5.1 · p. 7, §2.5.2 · p. 8
2.5.3. Inter-Robot Separation
- Six inter-robot safety formulations compare hard or slacked barriers, with the last three pairing a slacked barrier anchor with a body, forward-handle, or two-step rate condition.The body anchor protects the footprint at every stage; the look-ahead variant evaluates the rate condition on a forward handle, while the higher-order variant uses successive discrete-time barrier values.Source: §2.5.3 · pp. 8–9
- A hard barrier forbids predicted configurations inside the safe distance but can make a robot's OCP infeasible when several neighbors surround it.The hard discrete-time CBF limits how quickly the barrier may shrink, prompting earlier slowing, but retains the hard-constraint infeasibility risk in dense congregations.Source: §2.5.3 · p. 8
- The look-ahead rate condition acts on a forward handle that brings turn rate into the one-step barrier, allowing steering around a neighbor while the body anchor guards the footprint.The handle is placed a fixed distance ahead, with that distance less than the safe clearance.Source: §2.5.3 · p. 9
2.5.4. Obstacle Avoidance
- Obstacle avoidance applies the selected inter-robot safety formulation to fixed obstacle positions, using an obstacle-specific clearance and gains.At the terminal stage, all formulations revert to a plain terminal barrier.Source: §2.5.4 · p. 10
2.6. Look-Ahead Reference Point and Control Effort
The behavior module supplies a look-ahead reference that the per-robot OCP tracks with a quadratic cost. A separate control-effort term regularizes inputs for smoother, more energy-efficient trajectories.
2.6. Look-Ahead Reference Point and Control Effort
- The behavior module's look-ahead reference enters each robot's OCP through a quadratic tracking cost.The reference contains position, heading, linear speed, and angular velocity.Source: §2.6 · p. 10
- Heading deviation is wrapped to (−π, π] before it contributes to the reference-tracking cost.This avoids discontinuities at ±π.Source: §2.6 · p. 10
- A control-effort term regularizes the inputs to support smoother, more energy-efficient trajectories.The position, heading, speed, and angular-velocity tracking terms have positive weights.Source: §2.6 · p. 10
2.7. OCP Formulation
Each robot minimizes a finite-horizon OCP combining reference tracking, control effort, flocking costs, and safety-slack penalties. Dynamics, state and input bounds, and barrier-based safety constraints define the feasible problem.
2.7. OCP Formulation
- Each robot minimizes a finite-horizon OCP over N stages, subject to dynamics, state and input bounds, and inter-robot and obstacle safety constraints.The OCP incorporates neighbor data as stage-wise parameters.Source: §2.7 · pp. 10–11
- The stage cost combines look-ahead tracking, control effort, cohesion, alignment, and safety-slack penalties; separation and obstacle avoidance are enforced as constraints, not cost terms.The formulation's safety constraints are evaluated on neighbors' predicted trajectories and obstacle positions.Source: §2.7 · pp. 10–11
- Predictive CBF constraints requiring one- or two-step predictions are omitted at the terminal stage, where the formulations use a plain terminal barrier.Hard formulations fix their slack variables to zero.Source: §2.7 · p. 11
2.8. Simulation Workflow
At each control step, robots synchronously solve local OCPs using neighbor information from the previous timestep. Each applies its first action and publishes its updated predicted trajectory for neighboring robots.
2.8. Simulation Workflow
- Robots synchronously compute their control sequences from previous-timestep neighbor information, preventing use of another robot's within-step update.Each local OCP is solved over horizon N with a single SQP-RTI step.Source: §2.8 · p. 11
- A KD-Tree supports all in-range K-nearest-neighbor queries in O(N_b log N_b) total time, versus O(N_b^2) for naive pairwise search.The tree is constructed from current robot positions at the beginning of each control step.Source: §2.8 · p. 11
- Each robot samples local light intensity and gradient, gathers nearby obstacles and buffered neighbor trajectories, then applies its first control action and publishes its updated trajectory.The light sample determines a look-ahead reference according to the active swarm behavior mode.Source: §2.8 · p. 11
3. Simulation Results
Simulations compare six safety formulations under a common distributed predictive controller, then test the selected look-ahead formulation across tasks and swarm densities. Look-ahead safety gives the strongest overall phototaxis trade-off, while collisions rise sharply near geometric capacity and collective behavior persists.
3.1. Simulation Setup; 3.2. Capacity and Operating Density
- The 0.8 m × 0.8 m arena uses a 7 cm safety buffer, yielding a 137-robot geometric packing limit and a 100-robot operating point with headroom.The operating point leaves 37 robots of geometric headroom; runs use 10 seeds and average steady-state quantities over the final 300 seconds.Source: §3.1 · p. 12, §3.2 · p. 13, Table 1 · p. 13
3.3. Safety Formulation Comparison
- Across six formulations, LA Slack Dt-CBF gives the best phototaxis trade-off, with the widest high-density margins and fewest contacts.At Nb=137 it records 8.6 × 10^4 contacts versus 1.19 × 10^5 for Slack Dt-HOCBF; its per-robot solve time stays 3.66–3.74 ms across the sweep, within the 100 ms loop budget.Source: Figure 3 · p. 14, §3.3 · pp. 14–15
- Slacked formulations maintain at least 99.4% solver feasibility across densities, but collision counts distinguish them; anchors reduce contacts, and look-ahead improves the anchored variants.At Nb=100, anchored rate-condition formulations average 4.2 × 10^4 contacts versus 1.02 × 10^5 without an anchor; the look-ahead advantage over the chain grows to 1.39× at Nb=137.Source: §3.3 · pp. 14–15, Figure 3 · p. 14
3.4.1. Phototaxis Behavior
- With phototaxis, gradient following plus cohesion, alignment, and separation keeps the swarm compact and aligned; removing those costs and constraints collapses it onto the light peak.At large swarm sizes, convergent phototaxis and head-on encounters reduce local alignment, while nearest-neighbor distance falls to 5 cm near the packing limit.Source: §3.4.1 · pp. 16–18, Figure 5 · p. 17
3.4.2. Phototactic Orbital Behavior
- Phototactic orbiting converges from arbitrary starts to a compact, co-rotating formation, with more muted alignment loss than phototaxis and only mild obstacle effects.At Nb=137, nearest-neighbor distance is slightly below the 7 cm buffer; tangential motion aligns robots along their major axes.Source: Figure 5 · p. 17, §3.4.2 · p. 18, Figure 6 · p. 18
3.4.3. Flocking in the Absence of Stimuli
- Without a stimulus, cohesion and alignment form an ordered band along the arena walls, while crowding and avoidance steering increase heading disagreements.The band’s nearest-neighbor-distance trend resembles the phototactic orbital mode.Source: §3.4.3 · pp. 18–19, Figure 6 · p. 18
3.5. Collision Behavior Across Density and Tasks
- For deployed LA Slack Dt-CBF, contacts are negligible in small swarms but rise steeply toward packing capacity; obstacles matter most in small swarms, while task geometry changes collision load.Free space averages 1.0 contact at Nb=25, versus 8.6 × 10^4 at the packing limit; at Nb=100, phototaxis has 3.7 × 10^4 contacts and orbital behavior 3.5 × 10^2.Source: §3.5 · p. 19, Figure 7 · p. 19
3.6. Computational Performance and Scalability
- Per-robot solve-time mean and 99th percentile stay within 30 Hz budgets at every density, but the maximum observed solve of 98.8 ms meets 10 Hz, not 30 Hz.The single-process simulation’s per-step wall time grows with swarm size; the authors identify parallel execution as future implementation work.Source: §3.6 · p. 20
4. Conclusion
The controller unifies task-driven behavior, emergent flocking, and collision safety in a distributed NMPC architecture. The look-ahead Slack Dt-CBF performs best in collision prevention and inter-robot spacing, while the work remains limited by its kinematic robot model and internal safety comparison.
4. Conclusion
- Each robot solves local NMPC for task tracking, Reynolds-style flocking, and safety, exchanging predicted trajectories from active neighbors without a leader, central planner, or global map.The same architecture generalizes across phototaxis, phototactic orbital formation, and flocking without a stimulus.Source: §4 · p. 20
- The LA Slack Dt-CBF prevents collisions most effectively and provides the highest inter-robot margins for most tested swarm sizes.Pairing a slacked barrier anchor with a rate condition reduces contact counts by a factor of 2.4 versus formulations without an anchor; look-ahead gives the turn rate authority to steer around neighbors.Source: §4 · p. 20
- Orbital motion shows more muted order decay and more efficient packing than phototaxis, reflecting distinct circulating and convergent velocity-field structures in the paper's biomimetic interpretation.The authors suggest local polar-order and nearest-neighbor-distance metrics could connect these results to natural flocking research.Source: §4 · p. 20
- The bristlebot's vibration- and friction-driven locomotion is modeled kinematically, leaving surface, payload, and actuation asymmetries as unmodeled disturbances.The six safety formulations are compared within one controller stack, without external distributed-flocking baselines.Source: §4 · p. 21
Data availability
The paper reports no external data use and provides the code, simulation environment, and simulation videos online.
Data availability
- No external data was used; the code and simulation environment are available on GitHub, and simulation videos are archived on Zenodo.The archive includes full-length renders from which supplementary videos are derived.Source: Data availability · p. 21
A. Collective behaviors with static obstacles
Across arenas with two or twelve static obstacles, all four collective behaviors retain their qualitative form while routing around obstacles.
A. Collective behaviors with static obstacles
- All four modes retain the same qualitative behavior while routing the swarm around obstacles.This holds in arenas with two and twelve static obstacles.Source: A. Collective behaviors with static… · p. 25
- Figure 8 shows two static obstacles, while Figure 9 extends the scenes to twelve.Both use the LA Slack Dt-CBF formulation with N_b=50.Source: Figure 8 · p. 25, Figure 9 · p. 26
- Each figure presents four behavior rows across snapshots at 0, 225, 450, 675, and 900 seconds.The modes are phototaxis, phototaxis without Reynolds coupling, phototactic orbiting, and cohesion without a stimulus.Source: Figure 8 · p. 25, Figure 9 · p. 26
B. Supplementary data
Supplementary parameter information documents the experimental setup and controller settings used across the results.
B. Supplementary data
- Table 3 lists controller, safety, environment, and simulation parameters used for every result.The paper says this makes the Section 3 setup reproducible from the paper alone.Source: B. Supplementary data · p. 27, Table 3 · p. 27
- Cost weights, horizon, arena, and platform limits are shared across safety formulations, while selected CBF and slack parameters are retuned.The retuned parameters are α, L, γ1, γ2, and the slack weights.Source: Table 3 · p. 27
- The angular-rate bounds are asymmetric, reflecting the platform’s differing left- and right-turn authority.Source: Table 3 · p. 27