Current Version: v0.8.0
Bittle-only public release: the Petoi Bittle X runs in MuJoCo simulation and on real hardware over a WiFi bridge, with the SNN learning on-device via R-STDP. A CPG + SNN + cerebellum hybrid with an honest-claims framing throughout. The Go2 ablation (Paper 1) and Freenove sim-to-real (Paper 2) are preserved as paper tags.
Near Term
- Meta-Learning validation — longer runs (100k+ steps) and harder scenarios to activate the autonomous strategy adaptation loop
- Bittle hardware video — document the WiFi-bridge gait and on-device learning on the real Petoi Bittle X
- Bittle on-hardware learning — the SNN now learns directly on the Petoi Bittle X over WiFi; next is longer hardware runs and tighter sim-to-hardware agreement
- Resume training — validate brain persistence across sessions with meta-learning state
Medium Term
- Baby-KI — pure intrinsic reward (no external reward signal), autonomous learning through curiosity and prediction error
- Olfactory navigation — run-and-tumble chemotaxis with impulse-based orientation (biological E. coli strategy)
- Neuromorphic chip deployment — port SNN to Intel Loihi or SynSense Speck for sub-watt inference
- Conference submission — IROS, CoRL, or RSS paper with multi-platform results
Long Term
- On-hardware meta-learning — the real robot learns and adapts autonomously, not just replays sim-trained weights
- Multi-robot interaction — social behaviors between multiple MH-FLOCKE dogs
- Cortical layers — abstract planning, sequence learning, concept formation beyond the current cerebellar/brainstem level
Non-Goals
- LLM integration for motor control — language models are not motor controllers
- Scaling to millions of neurons without biological justification
- Cloud-dependent inference — MH-FLOCKE runs on-device, always