MORPH: Self-Organising Multi-Robot Task Allocation via Neuroplasticity-Inspired Adaptive Topology

AI summary

Overview

Research area: Multi-robot task allocation (MRTA), self-organising multi-agent systems, and neuroplasticity-inspired online learning, evaluated on a robotic warehouse benchmark.

Technical level: Advanced. The framework combines four biologically motivated plasticity rules (Hebbian learning, homeostatic scaling, structural synaptogenesis/pruning, and BCM-style metaplasticity) operating on a directed preference matrix; the motivation and high-level behaviour are accessible, but the update equations and parameter set require a background in multi-agent learning or computational neuroscience.

Scope: The paper introduces MORPH (Multi-agent Online Rewiring through Plasticity-guided Hierarchy), a training-free coordination layer that lets a heterogeneous robot fleet learn who should coordinate with whom online from task co-occurrence and delivery feedback, and evaluates it against all-to-all, proximity-based, and task-similarity baselines across fleet sizes of 8 to 24 agents.

What This Paper Is About

Multi-robot task allocation faces a tension: good coordination needs learned structure, but that structure must change when the environment changes. Existing methods sidestep this by assuming prior task knowledge at deployment time, such as a utility function, a cost matrix, or a policy trained offline, which makes them brittle when that knowledge is missing or when task demand patterns shift.

MORPH's goal is to remove those assumptions entirely. A directed pairwise preference matrix is initialised to zero at the start of each episode and updated only from runtime task co-occurrence and task-completion reward, so allocation preferences emerge online without a task model, bid computation, or offline training.

Key Contributions

1. MORPH as a training-free MRTA framework. The authors present what they describe as the first method in the single-task, single-robot, instantaneous-assignment class of the Gerkey–Mataric MRTA taxonomy to learn directed pairwise allocation preferences online, with a precise mapping of each of the four plasticity rules (synaptic, homeostatic, structural, metaplasticity) to the specific allocation failure mode it prevents.

2. Selective coordination beats dense coordination. Across all four tested fleet scales, MORPH exceeds all-to-all (Full Graph) throughput while using a progressively smaller fraction of possible coordination links as fleet size grows.

3. Mechanistic evidence that learned preferences are task-semantic, not spatial. The learned preference structure is reported to remain uncorrelated with Manhattan distance and to be driven by task co-occurrence history; the paper also reports two properties that emerge without being programmed: cross-type preference dominance and progressive preference sparsification.

4. An open implementation. Implementation and experiment scripts are released at https://github.com/Cyber-physical-Systems-Lab/morph_v2.

Main Findings

• Throughput advantage over all-to-all at every scale. MORPH reaches 110% of Full Graph throughput at N=24 (107.4 vs. 98.0 deliveries) while using only 21% of the possible links. At N=18 it reaches 102% of Full Graph throughput with 26% of links. At the smallest scale (N=8) MORPH matches Proximity at 62.4 deliveries (108% of Full Graph), using 27.0 of 28 possible links (97%).

• Link usage falls monotonically as the fleet grows. The active-link fraction drops from 97% at N=8 to 58%, 26%, and 21% at N=12, 18, and 24 respectively, meaning MORPH increasingly concentrates coordination on fewer relationships as the link budget grows quadratically.

• Proximity remains ahead at Medium and Large, but under a calibrated advantage. Proximity-r leads MORPH at N=18 (102.2 vs. 91.0 deliveries; p=0.008) and N=24 (113.2 vs. 107.4; p=0.038, Welch's two-sided t-test). The paper notes Proximity's radius was calibrated per scale to match MORPH's own converged link count, a procedure the authors state is unavailable at real deployment time.

• Task awareness alone is not enough. TSG, which uses the same co-assignment signal as MORPH but strips all four plasticity rules, achieves only 2–7 active links across scales and trails MORPH by 9.8 deliveries at N=24 (107.4 vs. 97.6).

• All four plasticity rules contribute. Removing neuromodulation produces the most consistent degradation across all four scales; removing the delivery reward gate causes an 8-point drop at Small (110% to 102%); removing the sliding threshold increases performance variance without a consistent mean-throughput penalty; and MORPH v1 (R1+R2+R3, no metaplasticity) underperforms the full system at all scales.

• Smaller degradation under repeated distribution shift. In the Medium-scale (N=18, 20 seeds) six-phase left/right alternating demand experiment, MORPH falls by 3.5 deliveries from Phase 3 to Phase 6 (−7.4%), versus −11.6 (−25.4%) for Proximity, −6.5 (−14.6%) for TSG, and −7.5 (−15.6%) for MORPH reset at each phase boundary. The abstract states MORPH degrades 3 times less than proximity-based methods; the body text reports approximately 3.3× versus Proximity and 2.1× versus resetting MORPH each phase.

• Recovery times shrink across repeated regimes for MORPH. MORPH's recovery time (steps to reach 90% of baseline delivery rate) drops from 36.9 at r3 to 17.7 at r6. The paper flags these recovery results as supporting evidence only, because variance is large.

• Persistent weights beat resetting over the long horizon. MORPH-reset relearns faster in earlier return phases (31.4 and 17.1 steps at r3 and r4), but ends Phase 6 at 40.8 deliveries versus 44.0 for MORPH, which the authors interpret as fast short-term relearning giving way to worse retention.

• Emergent structure without programming. A long-horizon episode (T=1500, Medium, N=18, seed=42) is used as mechanistic evidence, with W-matrix snapshots at t=0, 100, 200, 400, 600, and 800.

Methodology in Plain English

MORPH keeps a square table of numbers, one entry for each ordered pair of robots, recording how strongly robot i should request help from robot j. Every entry starts at zero each episode. The allocation policy converts each entry into a response probability, with a floor of 0.25 that provides residual coordination capacity before preferences are established, analogous to baseline synaptic transmission. MORPH sits on top of an existing controller: here, the Closest-Target Assignment heuristic with A* pathfinding, used unmodified.

Four rules update that table:

• Synaptic (R1, every step). A Hebbian update gated by two signals: how much two robots' assigned task sets overlap (a Jaccard index), and whether a delivery reward was just received. Only pairs whose Jaccard overlap exceeds 0.25 receive delivery-burst credit.
• Homeostatic (R2, every k=3 steps). Each agent maintains a correction term that biases link formation toward a target degree fraction of 0.35, preventing agents from becoming isolated or turning into hubs.
• Structural (R3, every k=3 steps). New links form when mutual information between two agents' task-assignment histories, plus a predictive hint and a homeostatic gain, exceeds an adjusted formation threshold; links are pruned when their weight falls below 0.008 and they are at least 20 steps old. The 20-step grace period protects nascent links from being depressed before they establish history.
• Metaplasticity (R4, every step and every k=3 steps). A per-link sliding threshold tracks each weight's history via an exponential moving average (τ=0.95) and adds extra decay to over-potentiated mature links. Separately, a global arousal signal compares the current delivery rate against a running baseline and shifts the system between consolidation (above expectation) and exploration (below expectation).

The rules deliberately run on two timescales (R1 and R4 every step; R2 and R3 every 3 steps), which the paper argues is functionally necessary: simultaneous weight and topology changes would destabilise the preference structure. The BCM threshold must also be updated before the Hebbian rule fires; updating it afterwards caused the two metaplasticity components to cancel each other out.

Evaluation uses TA-RWARE, a robotic warehouse benchmark with Automated Guided Vehicles transporting shelves to delivery stations and Picker agents assisting at pickup locations. Experiments span four fleet scales (Tiny N=8: 5 AGVs + 3 Pickers, 28 possible links; Small N=12: 8+4, 66 links; Medium N=18: 12+6, 153 links; Large N=24: 16+8, 276 links), 800 steps per episode, and 5 seeds, with all parameters held fixed across scales (Table 2: α=0.18, γ=0.98, β=0.04, θ_form 0.75→0.45, θ_prune=0.008, ρ_target=0.35, g=20 steps, k=3 steps, τ=0.95, α_r=0.08, neuromod gain/explore 0.4/0.10). Statistical comparisons use Welch's two-sided t-test with significance flagged at the 0.05 level. Three baselines represent different answers to the coordination-structure problem: Full-Graph (all-to-all), Proximity-r (pairs within a Manhattan distance radius, calibrated per scale to match MORPH's converged link count), and TSG (Task-Similarity Gate, which connects pairs sharing a current target with zero memory or plasticity).

Why This Matters

Impact on research. The paper argues that no existing MRTA method learns directed pairwise allocation preferences online from co-occurrence statistics alone while adapting continuously to change. MORPH reframes the allocation problem as a topology-learning problem and imports mechanisms from neural plasticity research, which has mostly been applied to weights within a single network (differentiable plasticity, Backpropamine, BCM-inspired weight adaptation), into a different object: the preference graph between agents. It also offers a concrete comparison point against self-organising approaches such as AMAS theory, stigmergy-based coordination, and response-threshold models, which it distinguishes by maintaining a directed, pairwise, continuously updated structure in agent memory rather than a scalar threshold or an environment-mediated signal.

Real-world applications:

• Robotic warehouses. AGV-to-Picker coordination, the exact setting of the TA-RWARE benchmark, where link budgets grow quadratically with fleet size and dense communication becomes noisy and expensive.
• Search-and-rescue operations. Teams whose task demand patterns shift as the operational area changes, which is the regime where Proximity and TSG degrade most.
• Autonomous vehicle fleets. Continuous reallocation as tasks arrive and complete, with no offline training phase before deployment.
• Systems running on existing controllers. Because MORPH is a preference layer requiring no modification to the underlying task policy, it can be added to an already-deployed fleet.

Industry relevance. The O(N²) preference lookups per step with no offline computation means the method is viable from the first episode in any environment, in contrast to learning-based allocation policies that require millions of environment steps on a fixed distribution or optimisation-based approaches that need a cost matrix and can require exponential-time solvers in the worst case. The finding that coordination can be concentrated on roughly a fifth of possible links at N=24 while beating all-to-all throughput is a direct communication-cost argument for large fleets.

Future Directions

• Close the gap with correctly specified spatial priors. Proximity still leads MORPH at Medium and Large scales when its radius is calibrated from MORPH's own converged link count. How to reach or exceed that performance without any spatial information is left open.
• Establish whether the pattern holds beyond these four scales. The study goes from N=8 to N=24; whether the monotonic link-sparsification trend continues at larger fleets is not reported.
• Strengthen the recovery-time evidence. The paper explicitly asks readers to treat recovery-time results as supporting evidence only because variance is large, indicating a need for more seeds or a different measurement design.
• Test transfer beyond the warehouse setting. The learned preferences are reported as task-semantic rather than spatial, but the evaluation is confined to one benchmark; the paper's discussion section, which positions MORPH in the MRTA design space and states its limitations, is referenced but its content is not included in the provided material.

Target Audience

Researchers and graduate students working on multi-robot task allocation, swarm and self-organising multi-agent systems, and biologically inspired learning for robotics. It will also interest warehouse-automation and fleet-coordination engineers evaluating whether a lightweight, training-free coordination layer can replace dense all-to-all communication or hand-tuned proximity heuristics in deployed systems. Readers need comfort with multi-agent coordination terminology and graph- or matrix-based formalisms; the neuroscience content is motivated rather than assumed.

Authors’ abstract

Multi-robot task allocation (MRTA) in dynamic environments faces a fundamental tension: effective coordination requires learned structure, but that structure must adapt when conditions change. Existing methods resolve this by assuming prior task knowledge, a utility function, a cost matrix, or a trained policy making them brittle when deployed without such knowledge or when task distributions shift. We present MORPH(Multi-agent Online Rewiring through Plasticity-guided Hierarchy), a training-free MRTA framework where global allocation quality emerges from 4 local plasticity rules (synaptic, homeostatic, structural, and metaplasticity) applied to a directed pairwise preference matrix updated from runtime co-occurrence and task-completion feedback. MORPH requires no task model, no bid computation, and no offline training; response decisions use learned AGV-to-Picker preferences rather than a fixed proximity rule. Within the Gerkey-Mataric MRTA taxonomy, MORPH is the first method in the single-task, single-robot, instantaneous-assignment class to learn directed pairwise allocation preferences online. Evaluated on the TA-RWARE warehouse benchmark (8-24 agents, 4 maps, 800 steps per episode, 5 seeds), MORPH achieves 110% of all-to-all throughput at N=24 while using only 21% of possible coordination links as an efficiency advantage that grows monotonically with fleet size. Under spatial task distribution shift, MORPH degrades 3x less than proximity-based methods while its learned preferences remain uncorrelated with Manhattan distance. Systematic ablation confirms all four plasticity rules contribute measurably. Two allocation properties emerge without programming: cross-type preference dominance and progressive preference sparsification, mirroring the developmental refinement of biological neural circuits. Learned preferences are driven by task co-occurrence history, not spatial proximity.