Reference
Version history
| Version | Change |
|---|---|
| 5.10.1 | A detector that cannot run raises instead of returning “nothing seen”; a watcher refuses to start rather than examine a hundred frames with a broken one. MediaPipe Tasks API supported alongside solutions. |
| 5.10.0 | Continuous local sensing: conditions a local sense can decide are watched at ~10 Hz instead of iterated through an LLM. |
| 5.9.1 – 5.9.5 | MediaPipe hands as a local geometric tier — 21 landmarks, finger counts, gestures — plus the routing fix that made it actually get called, and deterministic verdicts used directly rather than handed to a model to judge. |
| 5.9.0 | Vision-driven pursuit: a goal can watch until it sees something. |
| 5.8.1 – 5.8.6 | Capability gating (tools register only when the hardware answers a real probe), a [vision] extra that announces itself, the camera path on every platform, and describe_scene no longer reported as a fault when it succeeded. |
| 5.8.0 | SAM 3 open-vocabulary detection via a hosted GPU endpoint: pixel masks, centroids, bearings and cross-frame identity. |
| 5.7.0 | Vision through the existing provider abstraction — describe_scene on any vision-capable provider. |
| 5.6.x | Turn provenance, drive-rate calibration, proactive avoidance; stall detection that does not kill diagnosis. |
| 5.5.0 | Obstacle recovery, position held through failure, sensor power-sag diagnosis. |
| 5.4.x | Closed-loop goal pursuit for autonomous mode; AETHER_HOME for every store. |
| 5.3.0 | Angle-accurate turning, continuous operation, sensor-sweep rendering. |
| 5.0 – 5.2 | Multi-provider LLM abstraction with free tiers; capability gating, security hardening and the plan cache. |
| 4.5.x | Parallel autonomous execution with hard safety guards; reactive and observational sensing. |
| 4.3.0 | Vector retrieval over persistent memory, behind the [rag] extra. |
| 4.2.0 | Full-screen Textual TUI. |
| 4.1.x | Local analytics with opt-in telemetry, behind a consent gate on every entrypoint. |
| 4.0.0 | Autonomous mode: LLM goal decomposition and the execution loop. |
| 3.9.5 | Universal sensor-noise fix: drive_until_obstacle halts on consecutive bad reads, not a lifetime count — tolerates motor-EMI bursts. |
| 3.9.4 | Turn default 1.0 → 2.0 s; boot-time ultrasonic health probe + richer wiring guidance. |
| 3.9.3 | Continuous-motion primitives eliminate jitter; pulsed fallback for non-SunFounder adapters. |
| 3.9.1 | Latency skip on clean multi-step objectives; approach-speed curve. |
| 3.9.0 | Parallel execution layer; drive_until_obstacle skill. |
| 3.7.x | Auto-adapter system + the SunFounder safety chain. |
| 3.6.0 | MAVLink quadcopter integration with five arming rules. |
Roadmap
- ✓ Phase 1 — Reverse engineering: GPIO pin walk,
physical_map, LLM control. - ✓ Phase 2 — Robot Genome: versioned
robot.json, deterministic capabilities. - ✓ Phase 3 — Auto-adapter system: Tier 1/2/3; offline generation.
- ✓ Phase 4 — Vision-language grounding: four perception tiers (camera/VLM, YOLO, MediaPipe hands, SAM 3) and continuous local sensing.
- Phase 5 — Operational memory + sim-to-real action transfer.
- Phase 6 — Multi-robot coordination.