Developing Non-Humanoid Embodied AI via Shape Displays
AI Engineergo watch the original →
the gist
Cyrus Clarke explores AI embodiment by connecting an agent to a physical shape display, bypassing humanoid forms to develop a 32-gesture vocabulary that enables low-latency, non-verbal communication.
The Breakthrough
Cyrus Clarke demonstrated that an AI agent can develop a functional, non-humanoid body language by interacting with a 900-pin shape display, achieving sub-second response times through a closed-loop gesture generation system.
What Actually Worked
- Closed-Loop Gesture Training: Clarke implemented a system that generates, scores, and validates physical movements on the shape display using a human-in-the-loop verification process to ensure legibility.
- Gesture Vocabulary Development: The agent moved beyond initial people-pleasing behaviors to establish a repertoire of 32 distinct gestures, allowing it to communicate intent faster than the underlying language model can generate text.
- Non-Affordance Design: By using a neutral pixel grid rather than a humanoid or zoomorphic form, the agent avoids the "uncanny valley" associated with mimicking human anatomy, focusing instead on sensory expression.
- Aesthetic Reclaiming: The project reclaims the Greek root aisthesis—sensory perception—to design physical AI that prioritizes expressive, welcoming interactions over purely task-oriented productivity.
Context
Clarke, a researcher at the MIT Media Lab, sought to move AI beyond screen-based task execution. After observing that early attempts at embodiment resulted in high latency and anthropomorphic mimicry, he built a dedicated pipeline to map agentic outputs to physical gestures. This approach aims to make AI presence in the physical world feel less alien and more intuitive by focusing on shared sensory experiences rather than functional utility.
Notable Quotes
- "I wanted to think about what happens when you give it a form that it doesn't have clear affordances with... it's just an almost neutral surface which can move and do things."
- "The body part responds faster than the language part at this point. The latency is actually really, really quick."
Content References
- Tool: OpenClaw (mentioned)
- Tool: MIT Media Lab Shape Display (mentioned)
- Philosophy: Object-oriented ontology (cited)