D-MORPH: decentralised self-organising multi-robot task allocation via neuroplasticity-inspired adaptive topology

Xuezhi Niu and Didem Gürdür Broo · code: README · all numbers: results.md

Each robot in a heterogeneous team keeps its own directed partner preferences and links and updates them online with four plasticity rules (synaptic, homeostatic, structural and metaplastic) from its own observations: its target, the robot it is co-assigned with, and the completion of their joint mission. D-MORPH needs no task model, no bid computation, no message exchange and no offline training. It is evaluated on the TA-RWARE robotic warehouse with 8 to 24 robots.

102–108%of the throughput of a fully connected team, at four fleet sizes
9–23%of the possible robot pairs linked; the share falls as the fleet grows
−0.12 / −1.40deliveries per phase lost between repeated demand phases (Medium / Large), the least of all methods

Supplementary Videos 1–4

One episode (800 steps, seed 42) at each fleet size. Blue hexagons: AGVs; orange diamonds: Pickers; orange outlines: requested shelves; grey squares at the bottom: delivery stations. Green lines are established links (older than the 20-step grace period); red dashed lines are candidate links. The strip below shows cumulative deliveries and active links.

Video 1: Tiny, 5 AGVs + 3 Pickers (GIF)
Video 2: Small, 8 AGVs + 4 Pickers (GIF)
Video 3: Medium, 12 AGVs + 6 Pickers (GIF)
Video 4: Large, 16 AGVs + 8 Pickers (GIF)

Method

Allocation and learning at each step
Figure 1. Allocation and learning at each step. The benchmark takes the oldest pending task (1) and its nearest available robot i (2), and executes accepted tasks (5). Robot i requests the complementary robot j it prefers (3), which accepts according to its own preference, with probability 0.25 + 0.75 wj[i] (4). Each robot then applies R1–R4 to its own preferences and links (6).
RuleWhat each robot does with its own preferences and links
R1 synapticDecay, plus strengthening in proportion to the share of recent targets shared with the partner; completion credit when a jointly served mission is delivered.
R2 homeostaticAdjusts its readiness to form links so that its number of links stays near a target.
R3 structuralPrunes weak links after a 20-step grace period; forms a link to its best candidate (recent co-assignment and the robot it was proposed).
R4 metaplasticBCM-inspired sliding threshold (a preference above its own history decays faster) and neuromodulation from its own completed deliveries (learning rate and link-formation threshold).
The TA-RWARE warehouse with D-MORPH's links at four fleet sizes
Figure 2. TA-RWARE with D-MORPH's links at four fleet sizes: (A) initial state, before any task is allocated; (B) step 300. Green: established links; red dashed: candidate links.

Scale study

Seven methods, four fleet sizes, 20 seeds, 800 steps. Full Graph, MORPH, Proximity-r (r = 5) and TSG use MORPH's allocation loop; D-MORPH and both Random-Sparse variants use preference-based partner selection.

FleetD-MORPHFull GraphMORPHProximity-rTSGRandom-Sparse fixed / per-stepD-MORPH links (% of pairs)
Tiny (8)56.855.453.353.553.316.1 / 35.56.6 (23%)
Small (12)66.064.664.066.864.918.8 / 33.111.4 (17%)
Medium (18)90.784.288.385.486.212.4 / 47.117.6 (12%)
Large (24)104.399.2101.3104.2100.324.4 / 70.823.5 (9%)

Mean deliveries per episode. Paired two-sided Wilcoxon tests, p < 0.05 without correction for multiple comparisons: D-MORPH delivered significantly more than Full Graph at Medium (p = 0.0006) and Large (p = 0.0028) and more than MORPH at Medium (p = 0.0010); the differences at Tiny and Small are not significant.

Scale study: cumulative deliveries, throughput, active links and final deliveries
Figure 3. (A) Cumulative deliveries (mean ± SD). (B) Deliveries as a percentage of Full Graph; the y-axis is broken between the Random-Sparse variants and the other methods. (C) Active links as a percentage of the possible pairs. (D) Final deliveries (mean ± SD).
Deliveries relative to Full Graph versus active links, with the Proximity-r sweep
Figure 4. Deliveries relative to Full Graph against the number of active links (logarithmic x-axis). The blue curve is Proximity-r for r = 1 to 12; markers show the other methods.

Ablations

Each rule is removed in turn and compared with full D-MORPH on one measure fixed before the runs (docs/ablation_hypotheses.md), with the same 20 seeds. Completion credit, homeostasis, predictive formation and structural plasticity change their measures significantly; removing R4 (both parts) raises the number of active links by 2% at Medium and 4% at Large. Deliveries do not change significantly in any scale-study ablation; freezing the structure after step 400 lowers deliveries at Medium in the location shift (p = 0.022, uncorrected). Deliveries for every ablation are in docs/tables/tableS1_ablation_deliveries.tex.

Rule ablations: one pre-specified measure per rule, full D-MORPH against the rule removed
Figure 5. Full D-MORPH (filled circles) and D-MORPH with one rule removed (open diamonds), mean and 95% confidence interval over 20 seeds; text: paired difference (full minus ablation) and Wilcoxon p (bold: p < 0.05). (A) No completion credit (R1): preference of delivering pairs minus co-assigned-only pairs. (B) No homeostasis (R2): mean |links − target|. (C) No predictive formation (R3): active links at step 50. (D) Frozen structure (R3): co-assigned pairs that are linked after the first shift. (E) No R4 (BCM threshold and neuromodulation both off): mean active links. (A)–(D) are pre-specified; (E) was added afterwards.

Package location shift

New requests come only from the upper (U) or the lower (L) half of the shelves, alternating every 400 steps over six phases (2400 steps), at Medium and Large. The delivery stations are on the lower edge.

Delivery rate, deliveries per phase and paired differences under the package location shift
Figure 6. (A) Delivery rate (100-step moving average). (B) Deliveries per phase. (C) Degradation between repeated phases and (D) transient retention after a shift: D-MORPH minus each baseline, paired by seed (mean and 95% confidence interval; positive favours D-MORPH; circles: Medium, squares: Large). At Medium, the difference in degradation is significant against Full Graph (p = 0.033), MORPH (p = 0.007), Proximity-r (p = 0.007) and TSG (p = 0.021); at Large, no difference is significant (p ≥ 0.059).

Emergent allocation structure

Links and preferences over one long episode, links by pair type, link lifetimes
Figure 7. One 1500-step Medium episode. (A) Links and (B) preferences held by each robot (row i) at five snapshots; AGVs are robots 0–11. (C) Links by pair type. (D) Mean preference on links. (E) Lifetimes of pruned links: half end just after the grace period, about a third last more than 100 steps.

Reproduce

pip install -r requirements.txt
python -m experiments.run all --jobs 16   # results/*.json, one file per experiment
python -m experiments.analysis            # docs/figures, docs/tables, docs/results.md
python scripts/make_videos.py             # docs/videos

Every episode is deterministic given its seed. To run D-MORPH in another benchmark or simulator, see Using D-MORPH in your own benchmark in the README.

Cite

@unpublished{niu2026dmorph,
  author = {Niu, Xuezhi and G{\"u}rd{\"u}r Broo, Didem},
  title  = {{D-MORPH}: Decentralised Self-Organising Multi-Robot Task Allocation
            via Neuroplasticity-Inspired Adaptive Topology},
  note   = {Manuscript},
  year   = {2026}
}