DC-Ada: Reward-Only Decentralized Sensor Adaptation for Heterogeneous Multi-Robot Teams

Saad Alqithami

cs.ROApr 5, 2026

DC-Ada: Reward-Only Decentralized Sensor Adaptation for Heterogeneous Multi-Robot Teams

Authors: Saad Alqithami

Organizations: Department of Computer Science, Al-Baha University, Albaha 65779 Saudi Arabia

Abstract

Heterogeneity is a defining feature of deployed multi-robot teams: platforms often differ in sensing modalities, ranges, fields of view, and failure patterns. Controllers trained under nominal sensing can degrade sharply when deployed on robots with missing or mismatched sensors, even when the task and action interface are unchanged. We present DC-Ada, a reward-only decentralized adaptation method that keeps a pretrained shared policy frozen and instead adapts compact per-robot observation transforms to map heterogeneous sensing into a fixed inference interface. DC-Ada is gradient-free and communication-minimal: it uses budgeted accept/reject random search with short common-random-number rollouts under a strict step budget. We evaluate DC-Ada against four baselines in a deterministic 2D multi-robot simulator covering warehouse logistics, search and rescue, and collaborative mapping, across four heterogeneity regimes (H0--H3) and five seeds with a matched budget of $200{,}000$ joint environment steps per run. Results show that heterogeneity can substantially degrade a frozen shared policy and that no single mitigation dominates across all tasks and metrics. Observation normalization is strongest for reward robustness in warehouse logistics and competitive in search and rescue, while the frozen shared policy is strongest for reward in collaborative mapping. DC-Ada offers a useful complementary operating point: it improves completion most clearly in severe coverage-based mapping while requiring only scalar team returns and no policy fine-tuning or persistent communication. These results position DC-Ada as a practical deploy-time adaptation method for heterogeneous teams.

Figures & tables

Fig. 1: DC-Ada deployment and adaptation interface. Each robot adapts only a local observation transform

g_{\phi_{i}}

in front of a shared frozen policy

\pi_{\theta}

. Adaptation uses scalar rollout return broadcasts

R(\tau)

(once per rollout) and does not require sharing raw observations, maps, or gradients.

Task	Extra features $k$	Total dim $d$	Layout
Warehouse	5	73	$[p(2),\,v(2),\,k,\,\mathrm{lidar}(16),\,\mathrm{rgb}(32),\,\mathrm{depth}(16)]$
Search & Rescue	3	71	$[p(2),\,v(2),\,k,\,\mathrm{lidar}(16),\,\mathrm{rgb}(32),\,\mathrm{depth}(16)]$
Mapping	2	70	$[p(2),\,v(2),\,k,\,\mathrm{lidar}(16),\,\mathrm{rgb}(32),\,\mathrm{depth}(16)]$

TABLE I: Fixed-layout observation interface. All robots receive the same observation dimension within a task, regardless of sensor suite. Missing modalities are filled with neutral defaults. LiDAR rays are resampled to a fixed segment length of

16

Level	Robot	LiDAR	Range	Rays	RGB / Depth (FOV)
H0	R1–R4	yes	5.0	16	RGB+Depth ( $60^{\circ}$ )
H1	R1	yes	4.0	16	RGB+Depth ( $50^{\circ}$ )
	R2	yes	5.0	16	Depth only ( $60^{\circ}$ )
	R3	yes	6.0	16	RGB+Depth ( $70^{\circ}$ )
	R4	yes	4.0	16	Depth only ( $80^{\circ}$ )
H2	R1	yes	6.0	16	RGB only ( $60^{\circ}$ )

TABLE II: Sensor heterogeneity configurations (H0–H3) used to instantiate the four robots. When a modality is absent, the corresponding observation segment is filled with neutral defaults (zeros for RGB; ones for LiDAR/depth distance-like segments) to preserve a fixed observation layout across heterogeneity levels. Camera field-of-view (FOV) is

60^{\circ}

unless specified (H1).

Fig. 2: Merged environment/heterogeneity illustration: (a) severe heterogeneity rendered scene (conceptual), (b) sensor configuration examples motivating heterogeneity, and (c) top-down arena layout for the warehouse-style environment used in our simulator.

Hyperparameter	DC-Ada	Random Perturb.	Obs. Norm.	Local Fine-Tune	Shared Policy
Update interval $K$ (episodes)	3	3	–	1	–
Candidates $M$ (per update)	8	–	–	–	–
Noise scale $\sigma$	0.05	0.05	–	–	–
Step size $\alpha$	1.0	–	–	–	–
Short rollout fraction	0.25	–	–	0.25	–
Accept margin $\tau$	0.0	–	–	–	–

TABLE III: Key hyperparameters used in the final reported sweep. “–” indicates the parameter is not applicable. The shared policy is frozen during adaptation, and all methods respect the same matched interaction budget of

B{=}200{,}000

joint environment steps per run.

Task	Progress metric	Success criterion	Episode termination
Warehouse logistics	delivery ratio $=\text{delivered}/2$	delivered $\geq 2$	success or $T{=}500$ steps
Search and rescue	rescue ratio $=\text{rescued}/2$	rescued $\geq 2$	success or $T{=}500$ steps
Collaborative mapping	coverage	coverage $\geq 0.75$	success or $T{=}500$ steps

TABLE IV: Task specifications, termination, and success criteria used in the final reported sweep. Each evaluation run uses

N{=}4

robots, a horizon of

T{=}500

steps per episode, and a fixed interaction budget of

B{=}200{,}000

joint environment steps. The success criteria are deliberately non-trivial operational targets rather than one-event minima.

Task	Reward per step $r_{t}$
Warehouse	$r_{t}=-0.1+100\,\Delta\mathrm{delivered}+1\,\Delta\mathrm{picked\_up}-10\,\Delta\mathrm{collisions}+0.01\sum_{i=1}^{N}(20-d_{i})$ , where $d_{i}$ is the distance from robot $i$ to the nearest relevant target (package if not carrying; drop-off otherwise).
Search & Rescue	$r_{t}=-0.1+5\,\Delta\mathrm{found}+20\,\Delta\mathrm{rescued}+0.1\,\mathrm{health\_bonus}+0.5\,\mathrm{coverage}$ , where $\mathrm{coverage}$ is computed on a coarse grid and $\mathrm{health\_bonus}$ credits remaining victim health at rescue time.
Mapping	$r_{t}=-0.1+10\,\mathrm{coverage}+50\,\Delta\mathrm{coverage}+0.5\,\mathrm{spread}$ , where $\mathrm{coverage}$ is the explored-cell ratio and $\mathrm{spread}$ is the mean distance to the team centroid.

TABLE V: Reward functions used in the three domains (team-level scalar reward). All event terms use counter deltas to avoid reward explosion. Symbols follow the environment implementations.

Method	Mean episodes/run	Median runtime (s)	Median throughput (steps/s)	Scalar-feedback proxy (bytes/run)
Shared Policy	417.2	698.3	286.7	3338
DC-Ada	104.2	702.0	285.0	10693
Random Perturb.	425.0	727.7	275.1	3400
Local Fine-Tune	331.4	712.7	281.1	5302
Obs. Norm.	415.6	678.5	295.2	3325

TABLE VI: Runtime and scalar-feedback accounting for the final reported

B{=}200{,}000

-step sweep, aggregated over all environments, heterogeneity levels, and seeds. Runtime is measured wall-clock time per run on the evaluation hardware; throughput is median environment steps per second; the last column is an experiment-runner proxy that counts one 64-bit scalar team-return value per rollout. It should be interpreted as a low-bandwidth accounting convention rather than literal required deployment traffic for every baseline.

Environment	Method	H0	H1	H2	H3
Mapping	DC-Ada (Ours)	4757.4 $\pm$ 86.8	4760.2 $\pm$ 96.2	4717.6 $\pm$ 62.2	4721.8 $\pm$ 55.3
	Local Fine-Tuning	4763.7 $\pm$ 0.4	4765.4 $\pm$ 0.6	4758.2 $\pm$ 0.4	4781.3 $\pm$ 0.5
	Obs. Normalization	4715.1 $\pm$ 0.9	4710.0 $\pm$ 1.5	4684.6 $\pm$ 2.0	4694.6 $\pm$ 0.8
	Random Perturbation	4683.9 $\pm$ 51.3	4682.0 $\pm$ 53.0	4694.5 $\pm$ 49.0	4689.2 $\pm$ 58.1
	Shared Policy	4770.5 $\pm$ 0.3	4773.0 $\pm$ 0.3	4765.2 $\pm$ 0.8	4784.0 $\pm$ 0.8
Search & Rescue	DC-Ada (Ours)	14.1 $\pm$ 2.0	13.5 $\pm$ 1.1	15.2 $\pm$ 2.2	15.2 $\pm$ 1.0

TABLE VII: Performance comparison across heterogeneity levels under a fixed budget of

B{=}200{,}000

environment steps. Values are mean episode reward

\pm

std over 5 seeds.

Fig. 3: Scaling with heterogeneity level (H0–H3) for the three tasks. Curves show mean reward under a fixed interaction budget (

B{=}200{,}000

joint environment steps). A flatter slope indicates greater robustness to sensor mismatch for that task.

Fig. 4: Compact per-domain reward summary by method and heterogeneity level (H0–H3). Bars are averaged over five seeds under the same matched budget of

B{=}200{,}000

joint environment steps. This figure complements Fig. 3 by making within-level method ranking explicit.

Fig. 5: Success rate by method and heterogeneity level (H0–H3) using the thresholds in Table IV . Success is a binary completion metric; Sec. IX-F reports continuous progress metrics and threshold sensitivity under severe heterogeneity.

Metric	Contrast	Mean $\Delta$ [95% CI]
Reward	Shared Policy $-$ DC-Ada	$+62.2$ $[-15.1,\ 139.5]$
Success rate	DC-Ada $-$ Shared Policy	$+8.38$ pp $[6.75,\ 10.02]$ pp
Coverage	DC-Ada $-$ Shared Policy	$+0.0190$ $[0.0159,\ 0.0220]$

TABLE VIII: Selected seed-paired contrasts for collaborative mapping under severe heterogeneity (H3;

n{=}5

seeds). Positive values favor the first method named in the contrast. “pp” denotes percentage points.

Panel of Figure 14 — Fig. 6: Reward heatmaps across methods and heterogeneity levels. Uniform color across a row indicates robustness across H0–H3 for that environment.

Panel of Figure 15 — Fig. 7: Success-rate heatmaps across methods and heterogeneity levels using the thresholds in Table IV . Sec. IX-F contextualizes these values using continuous progress and threshold sensitivity under H3.

Environment	Method	Progress	Success (%)
Mapping	DC-Ada (Ours)	0.675 $\pm$ 0.002	12.4 $\pm$ 1.2
	Local Fine-Tuning	0.658 $\pm$ 0.000	4.5 $\pm$ 0.2
	Obs. Normalization	0.625 $\pm$ 0.000	0.0 $\pm$ 0.0
	Random Perturbation	0.659 $\pm$ 0.005	6.2 $\pm$ 1.8
	Shared Policy	0.656 $\pm$ 0.000	4.0 $\pm$ 0.0
Search & Rescue	DC-Ada (Ours)	0.302 $\pm$ 0.013	14.7 $\pm$ 1.1

TABLE IX: H3 progress and success under severe heterogeneity. Mean

\pm

std over 5 seeds. Progress is delivery ratio (Warehouse), rescue ratio (Search & Rescue), and coverage (Mapping). Success uses the thresholds in Table IV .

Fig. 8: Threshold sensitivity under severe heterogeneity (H3). Warehouse: delivered

\geq k

; Search-and-Rescue: rescued

\geq k

; Mapping: coverage

\geq\tau

. Curves indicate how completion probability varies with task strictness and contextualize rare-event success rates.

Environment	Variant	Reward	Success (%)	Progress	Episodes
Mapping	Shared policy	4784.0 $\pm$ 0.9	4.0 $\pm$ 0.0	0.656 $\pm$ 0.000	403.0 $\pm$ 0.0
	Full DC-Ada	4721.8 $\pm$ 61.9	12.4 $\pm$ 1.3	0.675 $\pm$ 0.003	102.0 $\pm$ 0.0
	No CRN	4775.9 $\pm$ 9.7	5.9 $\pm$ 1.2	0.661 $\pm$ 0.003	102.0 $\pm$ 0.0
	$M{=}1$	4804.6 $\pm$ 49.9	5.8 $\pm$ 1.2	0.659 $\pm$ 0.002	243.6 $\pm$ 0.9
	$M{=}4$	4771.8 $\pm$ 77.9	9.2 $\pm$ 1.9	0.672 $\pm$ 0.002	153.4 $\pm$ 0.9
	Always-best	4762.9 $\pm$ 100.3	9.6 $\pm$ 2.5	0.673 $\pm$ 0.004	102.0 $\pm$ 0.0

TABLE X: Targeted DC-Ada design ablations under severe heterogeneity (H3). All runs use the same frozen shared policy, five seeds, and a matched

200{,}000

-step budget. Mapping progress is mean coverage; search-and-rescue progress is mean rescue ratio. Values are mean

\pm

standard deviation across seeds.

Environment	Stress condition	Shared	DC-Ada	Local FT	Obs. norm.
Mapping	Clean	4.0 / 0.656	12.4 / 0.675	4.5 / 0.658	0.0 / 0.625
	Gaussian noise	4.0 / 0.656	11.4 / 0.675	4.5 / 0.658	0.0 / 0.633
	10% dropout	2.5 / 0.652	10.6 / 0.673	2.8 / 0.653	0.0 / 0.627
	2-step delay	4.0 / 0.656	10.8 / 0.674	4.5 / 0.658	0.0 / 0.625
	Combined mild	3.7 / 0.655	12.0 / 0.676	4.3 / 0.657	0.2 / 0.632
Search–Rescue	Clean	12.5 / 0.300	14.7 / 0.302	11.9 / 0.291	10.5 / 0.290

TABLE XI: Observation-stress evaluation under severe heterogeneity (H3). Each cell reports success rate (%) / progress, where progress is coverage for mapping and rescue ratio for search-and-rescue. Values are means over five seeds under the matched

200{,}000

-step budget.

Environment	Method	n	Reward	Success (%)	Progress
Warehouse	Shared policy	10	251.5 $\pm$ 0.1	1.2 $\pm$ 0.0	0.071 $\pm$ 0.000
	DC-Ada (Ours)	10	253.2 $\pm$ 1.7	1.4 $\pm$ 0.5	0.056 $\pm$ 0.011
	Random perturbation	10	234.0 $\pm$ 3.0	2.0 $\pm$ 0.3	0.084 $\pm$ 0.005
	Local fine-tuning	10	251.0 $\pm$ 0.3	1.3 $\pm$ 0.1	0.071 $\pm$ 0.002
	Obs. normalization	10	268.9 $\pm$ 0.1	1.5 $\pm$ 0.0	0.079 $\pm$ 0.001
Search–Rescue	Shared policy	10	16.05 $\pm$ 0.08	12.5 $\pm$ 0.1	0.299 $\pm$ 0.001

TABLE XII: Extended-seed confirmation for clean H3 across all three environments. This table combines the original five clean H3 seeds with five additional seeds (10 seeds total) for the five main methods. Progress is delivery ratio for warehouse, rescue ratio for search–rescue, and coverage for mapping. The main all-environment tables retain the uniform five-seed H0–H3 sweep; this targeted check evaluates whether the severe-heterogeneity trends persist under additional seeds. Values are mean

\pm

standard deviation across seeds.

Saad Alqithami received the M.Sc. (2012) and Ph.D. (2016) degrees in Computer Science from Southern Illinois University Carbondale, USA. He is currently an Associate Professor in the Department of Computer Science at Al-Baha University, Saudi Arabia. His research interests include artificial intelligence, agents & multi-agent systems, network science, social network analysis, and applied machine learning for resilient systems, with applications spanning healthcare, sustainability, and tourism. He has led funded research and innovation projects, and he actively serves the research community through editorial and peer-review activities.

Table 21

Explore similar work

Jul 6, 2026

Multi-Robot Open Adaptive Teaming Across Unseen Environments, Partners, and Scales

Deploying robot teams in the real world requires simultaneous adaptation to unseen environments, unknown partners, and varying team sizes, yet existing approaches often address these challenges in isolation under the closed-world assumption of fixed teammates. We formalize this as open adaptive multi-robot teaming and propose a hypergraphic-form game formulation that captures team-level cooperative relationships beyond pairwise interactions, providing a principled foundation for coordination structure inference when team composition changes dynamically within episodes. Unlike graph neural network architectures, this is a game-theoretic construct for modeling strategic interactions and payoff structures among agents. Building on this formulation, we develop the Hypergraphic Open-ended Learning Algorithm (HOLA), which progressively expands partner and environment diversity during training rather than optimizing for fixed configurations. Evaluated on cooperative pursuit with multi-drone and multi-quadruped platforms, HOLA outperforms all baselines across all three adaptability dimensions. Learned policies transfer directly to physical hardware without fine-tuning, with successful deployments on Crazyflie and Zsibot L1 platforms confirming robust real-world coordination in novel environments with unseen teammates.

Yang Li, Feng Xue, Fan Mo +6

1Shanghai Jiao Tong University · 2Sun Yat-sen University · 3National University of Singapore +4

Robot Policy Adaptation Multi-Robot Systems Multi-Robot Coordination Robot Policy Generalization Cooperative MARL

Ask PDF

Sep 26, 2026

Residual Denoising Enables Sample-Efficient Multi-Agent Coordination on Demand

Pretrained robot policies offer strong manipulation skills but are typically limited to single-agent settings, where a robot acts in isolation. In this work, we study how to adapt pretrained single-agent diffusion policies to multi-agent settings using minimal collaborative data, co-optimizing for two key objectives: high coordination performance and single-agent skill retention. To this end, we introduce ALTER, an adaptation method for coordination on demand: the adapted policy coordinates with other robots when deployed in a team while remaining capable of acting independently when operating alone. Execution is decentralized: each robot acts only on its own visual observations, without explicit inter-agent communication. Our method trains a coordination head that predicts a residual denoiser to transform single-agent behavior into coordinated multi-agent behavior when necessary while also preserving single-agent capabilities. To preserve single-agent capabilities, we augment a small number of collaborative demonstrations with self-distilled data generated by the base policy during training of the residual denoiser. In simulation, ALTER achieves higher coordination success over our baselines while retaining much higher source-skill retention. In our hardware experiments, we find similar trends where ALTER better co-optimizes for coordination success and single-agent skill retention than the baselines.

Dayi Dong, Maulik Bhatt, Aayushi Shrivastava +2

Department of Mechanical Engineering, University of California Berkeley, Berkeley, CA 94709, USA

Multi-Agent Coordination Robot Policy Adaptation Diffusion Policy Multi-Robot Coordination Data-Efficient Robot Learning Residual Policy Learning

Ask PDF

Jun 10, 2026

CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy

Multi-robot collaboration allows robots to efficiently take on a wide range of tasks, from moving a couch through a doorway to assembling structures on a construction site. However, achieving such coordination in mobile multi-robot settings remains challenging: centralized methods conditioned on the combined observations of a team scale poorly with team size, and decentralized methods that train one policy per robot often require explicit alignment procedures or information sharing at inference time to overcome partial observability. Our key insight is that the visuomotor priors of pretrained vision-language-action (VLA) models should enable reactive, decentralized collaboration from each robot's local observations alone, without these inference-time assumptions. We propose CHORUS, a framework that adapts a single VLA backbone to control diverse, multi-robot teams. At inference time, each robot runs an independent copy of CHORUS, conditioned only on its own observations and a robot-identifying prompt. In real-world experiments including mobile tape measurement, library book handovers, and laundry basket lifting, CHORUS achieves a 64% point improvement over decentralized, from-scratch models, improves reactivity to teammate behavior by 40% points, and outperforms centralized baselines. Together, these results show that a shared VLA backbone is capable of achieving decentralized multi-robot collaboration, without per-robot policies or inter-robot communication at inference.

Ria Doshi, Tian Gao, Annie Chen +2

Stanford University

Mobile Robotics Cross-Embodiment Robot Learning Multi-Robot Coordination Vision-Language-Action Models

Ask PDF