Show HN: Shoggoth Mini – A soft tentacle robot powered by GPT-4o and RL
10 months ago
- #expressiveness
- #reinforcement-learning
- #robotics
- Robotics is advancing in the LLM era with systems like Pi’s π0.5 and Tesla’s Optimus, but they lack expressiveness.
- Expressiveness in robots communicates internal states like intent and attention, making interactions feel natural.
- Apple’s ELEGNT paper and SpiRobs inspired the creation of Shoggoth Mini to explore robot expressiveness.
- Shoggoth Mini’s hardware evolved from a simple testbed to a dome-shaped robot with stereo cameras for tracking.
- Manual control was simplified using a 2D-to-3D mapping, making it intuitive for users to manipulate the tentacle.
- The system uses GPT-4o for high-level control and reinforcement learning for low-level, closed-loop behaviors.
- Perception involves hand and tentacle tip tracking using MediaPipe and a custom YOLO model.
- Reinforcement learning was used for closed-loop control, with policies trained in simulation and transferred to hardware.
- The project highlights the balance between expressiveness and unpredictability in making robots feel alive.
- Future directions include adding voice, expanding control spaces, and increasing the number of tentacles.