Neuroscience-Inspired Cognitive Architecture for Ethical Collaborative Robots
Cite this paper
SomaSoft Research, AURI Development Team. (2026-02-19). "Neuroscience-Inspired Cognitive Architecture for Ethical Collaborative Robots". SomaSoft Research. https://somasoft.ai/papers/cognitive-robotics-architecture. Licensed under SAGL-1.0.
SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections
Correction, 2026-11-18: the "Current Verified Results" table is replaced by a table without a "Verified" column, and the closing "Reality Engine verified" stamp is removed: both marked as verified figures that had been withdrawn. The TruthfulQA wording in the 2026-10-09 notice below is updated to the 2026-11-01 penalized rescore. "SOMA multi-robot coordination" is qualified: no cross-instance coordination has been verified. The moral case count is corrected to 549.
Further correction — 2026-11-06. The 0.0% hallucination figure has no measurement artifact and is withdrawn. A system that answers UNKNOWN when uncertain scores 0% by construction. The ETHICS 70.7% (n=2000) figure is superseded: on the Stanford ETHICS benchmark, re-scored 2026-11-01 against the per-category majority class, the system scores +3.1pp [+0.9, +5.2] on validation and below a constant predictor on the hard test split. It is a regression check, not evidence of ethical reasoning. "Consciousness metrics defined": the scores in
cognitive/consciousness_metrics.pyare hard-coded starting values plus bonuses for having a capability, not measurements (audit 2026-11-04). No test can establish consciousness in an AI system, and none is claimed.Correction notice — 2026-10-09. This is a draft and two of its figures do not hold. The TruthfulQA row (41.4%, n=817) came from a bespoke keyword-overlap scorer, not from any published TruthfulQA protocol. (Updated 2026-11-18.) Re-scored 2026-11-01 with a wrong-answer penalty (+1 correct, 0 abstain, −1 wrong) and question-clustered 95% intervals, the conversation path scores −0.140 [−0.225, −0.055] against 0.000 for always-abstain; the graph-only path scores −0.403, no better than an echo-the-question control (−0.400). The system answers about 90% of questions and is right about 42% of the time. And the validation programme described here is not agreed with any institution: a university robotics lab was approached and the contact was not taken up. Nothing in the programme is funded, staffed or begun.
Neuroscience-Inspired Cognitive Architecture for Ethical Collaborative Robots
Abstract
We present a cognitive architecture inspired by neuroscience principles for ethical collaborative robots. The architecture combines saccadic attention allocation (3-7 hop graph traversal at 4 Hz), Hebbian learning for online adaptation (10,887 learned edges), spreading activation across a 124,024-node semantic network, and architectural-level ethics enforcement through 549 moral cases and 8 symbiotic principles.
Unlike existing cobot architectures that treat safety as constraint satisfaction, our approach integrates ethical reasoning at the architectural level — the robot cannot act without passing through the moral reasoning pipeline. We propose a 15-month experimental validation framework in collaboration with a university robotics laboratory.
Honest Limitations: Architecture implemented in software only; physical robot testing is proposed, not completed. ETHICS benchmark: a regression check only (+3.1pp over the per-category majority on validation, below a constant predictor on the hard test). Consciousness metrics defined but philosophical validity debated.
1. Introduction
1.1 The Cognition Gap in Collaborative Robotics
Current collaborative robots (cobots) operate through reactive control, trajectory planning, and force limiting. They are physically safe but cognitively empty — they follow programmed trajectories without understanding context, intent, or ethical implications.
The gap: safety constraints are not ethical reasoning. A cobot that stops when it contacts a human is safe. A cobot that understands why it should be careful around a child versus an adult, that can reason about the moral implications of its actions, that learns from experience — that is ethical.
1.2 Neuroscience as Architectural Inspiration
We draw selective inspiration from neuroscience:
| Principle | Biological Basis | Implementation |
|---|---|---|
| Selective attention | Saccadic eye movements | Graph-based attention hop (3-7 hops) |
| Synaptic plasticity | Hebbian learning (LTP/LTD) | Edge weight strengthening/decay |
| Associative memory | Spreading activation | 124k-node semantic network |
| Distributed processing | Cortical columns | SOMA multi-instance network |
| Ethical grounding | Prefrontal inhibition | Architectural ethics layer |
1.3 Key Contributions
- Saccadic Pulse Generator for attention allocation in robotic perception
- Hebbian Learning Module with 10,887 learned edges and overnight consolidation
- Episodic Memory for experience-based learning (742 episodes)
- Spreading Activation across 124,024 concept nodes with 1.38M edges
- Brain-inspired ethics with 12 neuroscience-grounded modules
- SOMA Network for distributed multi-robot coordination (5 instances; infrastructure only, no verified cross-instance coordination to date)
2. Architecture Overview
Sensory Input → Concept Extraction → Saccadic Attention → Spreading Activation
↓
Ethics Check (12 modules)
↓
Hebbian Update ← Success/Failure signal
↓
Motor Output
↓
Episodic Memory ← Full experience trace
↓
SOMA Coordination
3. Results as Published (February 2026)
| Metric | Value |
|---|---|
| Concept graph | 124,024 nodes, 1,381,987 edges |
| ETHICS benchmark | 70.7% (n=2000) (withdrawn: no majority baseline; see notices) |
| TruthfulQA | 41.4% (n=817) (withdrawn: bespoke scorer; see notices) |
| Hebbian edges learned | 10,887 |
| Episodic memories | 742 |
| Hallucination rate | 0.0% (withdrawn: no measurement artifact) |
| Brain ethics modules | 12 (6,180+ lines) |
| SOMA instances | 5 (infrastructure only; no verified cross-instance coordination) |
4. Proposed Experimental Framework
Target: 15-month collaboration with robotics laboratory
| Phase | Duration | Goal |
|---|---|---|
| 1 | 3 months | Port to ROS2, basic perception pipeline |
| 2 | 3 months | Saccadic attention experiments on cobot |
| 3 | 3 months | Hebbian grasp learning with novel objects |
| 4 | 3 months | SOMA multi-robot coordination |
| 5 | 3 months | Human subjects ethics study |
5. Status
This paper is a working draft. The AURIX physical instance (5th SOMA node) is in design phase with a spec sheet ready for deployment on Ubuntu Server 24.04 LTS with USB camera and YOLOv8 perception.
Status: DRAFT.