Model card: point to experimental fast-brain07
Browse files
README.md
CHANGED
|
@@ -21,7 +21,9 @@ A **30,000-neuron recurrent network** built from a subgraph of the fruit-fly mal
|
|
| 21 |
|
| 22 |
This repository contains controller checkpoints, gate detectors and the derived flight graph. The controller receives telemetry-derived sensory inputs. For flights guided by camera images, a separate gate detector and hand-engineered pilot provide the goal it follows. Training uses imitation and back-propagation through a differentiable simulator; the Hub's reinforcement-learning label is a broad task category.
|
| 23 |
|
| 24 |
-
**Newest: [fast-
|
|
|
|
|
|
|
| 25 |
|
| 26 |
**For the earlier experimental visual stack, use the self-contained [scene09](https://huggingface.co/Skulitom/haltere/tree/main/scene09) or [motor10](https://huggingface.co/Skulitom/haltere/tree/main/motor10) bundles and their versioned instructions.** Motor10 candidate05 is fast-brain06's parent, not a promoted replacement for scene09. Its release evaluation recorded one generated-loop finish in four development attempts, including flag and pillar impacts on Straw Bale and Minus Two and a separate telemetry stop. The successful repeat took 3:01.576. No qualifying unseen-course or freestyle result is claimed. See the [motor10 provenance and limits](https://github.com/skulitom/haltere/blob/main/docs/motor_brain_10_release.md) and [continuing flight evaluation](https://github.com/skulitom/haltere/blob/main/docs/flight_cards/README.md).
|
| 27 |
|
|
|
|
| 21 |
|
| 22 |
This repository contains controller checkpoints, gate detectors and the derived flight graph. The controller receives telemetry-derived sensory inputs. For flights guided by camera images, a separate gate detector and hand-engineered pilot provide the goal it follows. Training uses imitation and back-propagation through a differentiable simulator; the Hub's reinforcement-learning label is a broad task category.
|
| 23 |
|
| 24 |
+
**Newest: [fast-brain07](https://huggingface.co/Skulitom/haltere/tree/main/fast-brain07)** (experimental): re-distilled under the arc-turn pilot, it finished Straw Bale in 5:45.792 and 5:46.846 and is smoother after checkpoints; Minus Two and Pine Valley not finished. [Release notes](https://github.com/skulitom/haltere/blob/main/docs/fast_brain_07_release.md).
|
| 25 |
+
|
| 26 |
+
**Previous: [fast-brain06](https://huggingface.co/Skulitom/haltere/tree/main/fast-brain06)** (experimental). With brain motors under the disclosed fast race-cue pilot at 6 m/s it finished the full Straw Bale race twice (5:50.489, 5:50.204; matched fast PD 5:02.933) but not Minus Two or Pine Valley, where obstacles beside or on the line to the checkpoint also stop the PD. Only readout rows 0-2 changed; scene currents are blanked. Seen development courses only. See the [release notes and limits](https://github.com/skulitom/haltere/blob/main/docs/fast_brain_06_release.md).
|
| 27 |
|
| 28 |
**For the earlier experimental visual stack, use the self-contained [scene09](https://huggingface.co/Skulitom/haltere/tree/main/scene09) or [motor10](https://huggingface.co/Skulitom/haltere/tree/main/motor10) bundles and their versioned instructions.** Motor10 candidate05 is fast-brain06's parent, not a promoted replacement for scene09. Its release evaluation recorded one generated-loop finish in four development attempts, including flag and pillar impacts on Straw Bale and Minus Two and a separate telemetry stop. The successful repeat took 3:01.576. No qualifying unseen-course or freestyle result is claimed. See the [motor10 provenance and limits](https://github.com/skulitom/haltere/blob/main/docs/motor_brain_10_release.md) and [continuing flight evaluation](https://github.com/skulitom/haltere/blob/main/docs/flight_cards/README.md).
|
| 29 |
|