Where the theory meets physical systems. The experiments, the instrumentation, and the engineering that turn the rest of the programme into something that moves in the world.
Robotics platforms here are treated as test instruments for specific theoretical claims, not as ends in themselves. A claim about delay-robust control is only as good as its behaviour on actual hardware; a claim about semantic mapping is only as good as the quality of the scene graphs it produces under realistic frame rates; a claim about reinforcement-learned assembly is only as good as the sim-to-real gap closes.
Three integrated pieces
1. Semantic-level perception — SEGO
SEGO (Semantic Graph Ontology mapper) unifies geometric SLAM, deep-learning object detection, and ontology-driven reasoning into a single pipeline that produces cognitive scene graphs consistent with both geometry and domain knowledge. On the TUM RGB-D dataset the SRQI (Semantic Recognition Quality Index) improves from 0.662 at 10 FPS to 0.703 at 30 FPS, then plateaus — aligning with the known limit of human perceptual integration.
See Cognitive Synergy Architecture: SEGO.
2. The five-pillar review
A survey of the architectural pieces needed for cognitive collaborative robots: semantic-level perception, cognitive action planning, explainable learning and control, safety-aware motion design, and multimodal human-intention recognition. The review proposes a unified Cognitive Synergy Architecture as the integration layer.
See Towards Cognitive Collaborative Robots.
3. Production-grade assembly — Soma cube on a Doosan M0609
The most hands-on piece in the collection. A 6-DOF collaborative robot learns to assemble the seven distinct Soma-cube pieces into a cube using legal-action masked DQN with safe ZYZ regrasp. Key numbers:
- Action space 4,536 → 2,484 feasible actions (26% reduction).
- Motion success rate 54 % → 96 % after adding the ZYZ regrasp.
- Sim-to-real 75 % assembly success with ±1.8 mm positioning.
- Whisper-based Korean-language HRI at 94 % recognition accuracy.
See Learning to Assemble the Soma Cube.
Papers on this track
- Cognitive Synergy Architecture: SEGO for Human-Centric Collaborative Robots
- Towards Cognitive Collaborative Robots
- Learning to Assemble the Soma Cube with Legal-Action Masked DQN and Safe ZYZ Regrasp on a Doosan M0609
Related
- Ontology Neural Networks — the representational layer that these systems would ideally consume.
- Delay-robust control — the stability guarantees that close the loop around the robot.