Controlling Collectives of AI Agents in Reasoning Space with Spatial Transformers
Frederic Vatnsdal, Roshan Gopal, Romina Garcia Camargo, Vijay Kumar, Alejandro Ribeiro
cs.RO · cs.AI
Quick verdict
COMPASS couples frozen small LMs with a spatial transformer for decentralized flocking; strong sim results to N=1024, but simulation-only with thin baselines.
Summary
The paper introduces COMPASS, a decentralized architecture for controlling large fleets of robots with natural-language commands. Each robot runs a frozen small language model (Qwen3-1.7B) whose context consists of an operator prompt, a generated thinking trace, a learned goal token, and a learned feedback token. A 4-layer multi-agent spatial transformer (MAST) with rotary positional encodings diffuses reasoning states hop-by-hop over the communication graph at 10 Hz, producing both low-level actions and the feedback token refreshed at 1 Hz, so the LM never processes raw neighbor states. Trained by behavior cloning plus DAgger to imitate a centralized potential-field expert on teams of 32-64 robots, COMPASS evaluates zero-shot up to N=1024, with heading error falling from 2.79 degrees at N=8 to 0.34 degrees at N=1024. Ablations show language-only communication fragments a 64-robot flock into 15 groups, while a centralized Claude Opus 4.8 controller degrades as team size grows across an 8-128 robot sweep. Zero-shot generalization holds for unseen prompt templates (median 1.9 degrees), conceptual commands such as 'where Santa Claus lives' (5.7 degrees), and secondary intercardinal directions (mean 7.96 plus/minus 4.41 degrees, NNE failing at 13.2 degrees). Structured prompt diversity helps only when an easier phrasing is mixed in. Main limitations: simulation-only results (double integrators, ideal communication), no learned baseline without the LM, thin statistics (small trial counts, mixed mean/median, no confidence intervals), and no code release stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
6 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- COMPASS generalizes zero-shot to fleets up to 16x its training scale, flying N=1024 simulated robots with heading error decreasing from 2.79 deg (N=8) to 0.34 deg (N=1024).
- A compact learned feedback token in the LM's latent space is essential: routing raw neighbor state through the language channel fragments a 64-robot flock into 15 groups, performing worse than no communication at all.
- The learned goal-token interface grounds unseen language zero-shot: held-out templates (1.9 deg median), conceptual commands (5.7 deg), secondary intercardinals (7.96+/-4.41 deg mean), and numeric bearings (2.2+/-1.6 deg mean) are flown without retraining.
- Structured diversity helps only when an easier phrasing is mixed into the fleet (1/4 plain), which the authors attribute to LM bias cancellation; equal-difficulty diversity yields no gain.
- A centralized frontier LLM (Claude Opus 4.8) decodes directives accurately but cannot coordinate per-robot accelerations at scale, degrading across the 8-128 robot sweep (Fig. 4).
Strengths
Zero-shot scale generalization with error improving under scale
Trained only on N=[32,64] but evaluated to N=1024 (16x), with heading error falling from 2.79 deg at N=8 to 0.34 deg at N=1024 (Sec. IV-A, Fig. 3); translation-invariant 2D rotary encodings and graph-masked attention (Sec. II-B) make the same parameters valid across scales.
Informative negative ablations isolating the communication channel
Language-only communication (raw neighbor state in the prompt) is worse than no communication: the N=64 flock fragments into 15 groups at 2.5 m/s, half the target speed (Sec. IV-A); this concretely supports the learned-latent-token design over language-channel feedback.
Modular frozen-LM design with quantified engine tradeoffs
Broad, honestly-graded generalization taxonomy
Non-obvious structured-diversity finding
Complete architectural and optimization specification
Weaknesses
Simulation-only validation with idealized assumptions
Missing control: learned prompt conditioning without the LM
Thin and inconsistent statistics
Under-specified centralized LLM baseline
Internal bookkeeping inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
COMPASS (Sections I-II) is a decentralized architecture for language-commanded flocking that couples a frozen small language model (Qwen3-1.7B) on every robot with a 4-layer multi-agent spatial transformer (MAST). Each robot's LM context is a persistent program of embedded tokens - operator prompt P, a generated thinking trace T, a learned feedback token f, and a learned goal token g (Eq. 1); a latent reasoning state z_i is read at the goal-token position after the first l layers (Sec. II-A.2). MAST diffuses these states hop-by-hop over the communication graph (rc = 4 m, 2D rotary encodings with period 96 m) at 10 Hz, emitting both the low-level action and the feedback token refreshed at 1 Hz, so the LM never sees raw neighbor state. The headline claims: zero-shot scale generalization to 16x the training fleet (N=1024 vs trained 32-64) with heading error falling from 2.79 deg (N=8) to 0.34 deg (N=1024) (Sec. IV-A); the necessity of a learned latent feedback token - language-only communication fragments a 64-robot flock into 15 groups and is worse than no communication; and zero-shot generalization to unseen templates, conceptual commands, secondary intercardinals, and numeric bearings.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 6 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Is COMPASS tested on real robots?
No. All reported results are in simulation with double-integrator dynamics and an idealized 4 m communication graph without latency or packet loss. The '1024 robots' in the abstract are simulated agents, not physical drones.
What language model does COMPASS use?
Qwen3-1.7B, frozen (not fine-tuned), running locally on each robot. The paper also probes Qwen3-0.6B and 4B as swappable reasoning engines, finding larger models reduce heading error on vague commands but cost more latency and memory (Table II).
How many robots can COMPASS control?
It is trained on 32-64 robot teams and evaluated zero-shot up to 1024 simulated robots, where mean heading error was lowest (0.34 degrees); error decreased monotonically with team size across the 8-1024 sweep.
What is the learned feedback token in COMPASS?
It is a compact latent vector produced by the multi-agent spatial transformer from multi-hop neighbor information and injected into each robot's language model context. Its purpose is to let the LM incorporate fleet state without raw numbers in the prompt - the paper shows that putting raw neighbor states in the language channel instead makes the flock fragment.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as extracted captions and axis labels; curve-level data (e.g., the log-scale sweeps in Fig. 3 and retry/latency curves in Fig. 4) cannot be independently verified from the text.
- Review is based on arXiv v1 (2609.28247, 23 Sep 2026) single-column extraction; any appendix or supplementary material is not included.
- Some procedural details (DAgger iteration counts, expert gain values, centralized-LLM prompt budgets) may exist in portions of the paper that were truncated during extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.