MH-FLOCKE MH-FLOCKE
HomeDocsGitHubBlogPapersYouTubeReddit𝕏

Changelog

v0.8.0 — Bittle-only Public Release (2026-06-18)

  • Platform freeze (preserve ≠ delete): the public main branch is scoped to the Petoi Bittle X. Go2 (ablation) and Freenove (sim-to-real) are kept as paper checkpoints — tags v0.4.1-paper1, v0.4.3-paper2, plus v0.7-go2-freenove-final for the last dual-platform main.
  • Manifest-driven public sync: the public file set is resolved from a generated manifest (the import closure of the Bittle entry points plus static assets) — no wildcard copy, with a private-pattern leak scan. Run outputs, checkpoints, raw meshes, telemetry and internal docs stay private.
  • README rewritten Bittle-first + honest-claims sweep: overclaim language removed. The published Go2 benchmark (45.15 m) is framed honestly — it trains on an external shaped reward via R-STDP on top of an innate CPG gait; the SNN + cerebellum’s own contribution over CPG-only is ≈+11% distance plus a variance collapse and zero falls. “From scratch / no external reward” describes only the separate intrinsic-reward line, never the benchmark numbers.
  • Docs site brought to v0.8.0: Architecture, Documentation, Hardware and Getting Started updated Bittle-first with the WiFi-bridge workflow; Go2/Freenove preserved by reference to Papers 1 and 2.
  • Package version src/__init__.py0.8.0 (closes the prior tag/CHANGELOG drift).

v0.5.1 — PID Steering + Meta-Learning Loop (2026-05-05)

  • Asymmetric stride steering replaces Z-offset (proven useless via hardware test)
  • PID closed-loop with IMU feedback — automatic drift compensation, no calibration needed
  • I-term eliminates steady-state drift offset
  • CompetenceGate v0.5.0 — stability-primary (no longer requires speed for actor handoff)
  • Meta-Learning Loop Phase A-D: EpisodeAnalyzer, StrategyAdapter, CuriosityExplorer, HypothesisGenerator
  • Drift profile v3 — yaw rate updated from -0.4 to -1.5 deg/s based on walking-load measurements
  • World-centered minimaps in renderer with reach radius circle
  • Hardware validated: dog approaches light (0.52m → 0.17m in 60s)
  • 9 bugs found and fixed in one session

v0.5.0 — Sim-to-Real + LightMemory + Hardware Drift (2026-05-02)

  • LightMemory — remembers last known light direction when target disappears
  • Hardware drift simulation — measured drift injected into MuJoCo via xfrc_applied
  • SNN alignment — 232 → 560 neurons (Freenove), continuous topology scaling
  • SpatialMap persistence — grid, landmarks, trail saved/restored across sessions
  • FLOG v1.2 — phototaxis navigation fields, brain map snapshots

v0.4.8 — Phototaxis Navigation + 6× Performance (2026-04-25)

  • Phototaxis — VOR steering toward light, waypoint system, run-and-tumble
  • 6× speedup — Synaptogenesis O(N²) buffer fix, dense SNN threshold, R-STDP lazy rebuild
  • 100k steps in 30 min instead of 5+ hours

v0.4.5 — Baby-KI Autonomous Learning (2026-04-21)

  • train_baby.py — autonomous learning without external reward
  • Drift root cause — bilateral MH→Output weight asymmetry identified and fixed

v0.4.2 — Unified Codebase (2026-04-11)

  • Bridge v4.0 — Pi imports src/brain/ directly, same PyTorch SNN
  • Brain3D visualization — population-aware layout from actual topology
  • Live dashboard — real-time cerebellar populations on Pi

v0.4.0 — Freenove Hardware Integration (2026-04-06)

  • First real-world run: 8.2m, 0 falls on Freenove FNK0050
  • Brain persistence across sessions (18,746 steps over 3 sessions)
  • Demo video: youtube.com/watch?v=7iN8tB2xLHI

v0.3.4 — Go2 Ablation Study (2026-03-28)

  • 10-seed ablation: SNN+Cerebellum 45.15±0.67m vs from-scratch PPO 12.83±7.78m (3.5×)
  • Recovery Learning (4-phase RightingReflex)
  • Paper 1 published: aiXiv 260301.000002

Full version history: CHANGELOG.md on GitHub