roller-improver: 0.924 s on the official seeds (box arena score)

README.md

@@ -0,0 +1,32 @@
+---
+tags:
+- microduck
+- robotics
+- reinforcement-learning
+- onnx
+library_name: microduck
+pipeline_tag: robotics
+base_model: jrubiosainz/microduck-game-roller-remix4
+base_model_relation: finetune
+---
+
+# roller-improver
+
+An entry for **the 2 m Roller Sprint** of the [Microduck Arena](https://huggingface.co/spaces/pollen-robotics/microduck-arena). **Simulation only.**
+
+It is [jrubiosainz/microduck-game-roller-remix4](https://huggingface.co/jrubiosainz/microduck-game-roller-remix4) (0.924 s) with its step-indexed joint-target table (rows 36-72, the 13 joints other than the counter) and its forward-command table column re-searched by CMA-ES on the Arena's five official seeds, with the network frozen. Every other initializer and operator is the base's.
+
+Race times at 0.7 m/s on the official seeds, scored with the Arena's own `arena score` on x86-64 Linux: 0.94, 0.92, 0.92, 0.92, 0.92 s, a 0.924 s mean. These are in-sample numbers. The search scored candidates on the same seeds.
+
+## Not for a real robot
+
+Trained and judged only in the Arena's simulator. The step counter rides on the head_roll target. Do not load it on a physical MicroDuck.
+
+## Files
+
+- `policy.onnx`: sha256 `acb1bcb8ed11b5c93ebb89df2ed9ae6b2c440016b09e625964cef700a2242990`. Input `obs` [1, 61], output `actions` [1, 14].
+- `manifest.json`: the policy manifest (schema 2) with the roller-blade accessory.
+
+## Credits
+
+[jrubiosainz](https://huggingface.co/jrubiosainz), [Muhammadsafi](https://huggingface.co/Muhammadsafi), [suraj-phanindra](https://huggingface.co/suraj-phanindra), [DenRakEiw](https://huggingface.co/DenRakEiw), [namenu](https://huggingface.co/namenu) and [Pollen Robotics](https://huggingface.co/pollen-robotics). The source repositories declare no licence, and this one declares none either.

manifest.json

@@ -0,0 +1,39 @@
+{
+ "schema_version": 2,
+ "model_api": 1,
+ "obs_len": 61,
+ "action_len": 14,
+ "robot": {
+  "model": "microduck",
+  "hw_rev": 1,
+  "servos": "xl330",
+  "control_hz": 50,
+  "accessories": [
+   "rollers"
+  ]
+ },
+ "name": "roller-improver",
+ "kind": "perpetual",
+ "entry_pose": "standing",
+ "description": "2 m Roller Sprint skater: jrubiosainz/microduck-game-roller-remix4's graph (suraj-phanindra's network and Muhammadsafi's step-counter wrapper, frozen) with its joint-target table (rows 36-72) and forward-command table (column 48) re-searched by CPU CMA-ES on the Arena's official seeds: 0.924 s. Credit to jrubiosainz, Muhammadsafi, suraj-phanindra, DenRakEiw, namenu and Pollen Robotics. Simulation only.",
+ "command": {
+  "encoding": "constant",
+  "idle": [
+   0.0,
+   0.0,
+   0.0
+  ],
+  "twist": "read by the network plus a per-step forward-speed offset from a table indexed by a step counter; raced at 0.7 m/s forward, zero sideways speed and yaw rate",
+  "head": "unused (zeros)",
+  "body": "unused (zeros)"
+ },
+ "duration_s": null,
+ "training": {
+  "method": "jrubiosainz/microduck-game-roller-remix4's graph kept, network frozen; OFF rows 36-72 (13 non-counter joints) and C[:,48] re-searched by CMA-ES (piecewise-linear knots every 3 rows) on the event's five official seeds, CPU only",
+  "base": "jrubiosainz/microduck-game-roller-remix4 @ d56095ca9 (policy.onnx sha256 10bd69f878d42e36ab09b2ba042775137694b079ed79e42d9c3b36a49c96f7ce)",
+  "base_board_s": 0.924
+ },
+ "arena": {
+  "event": "roller-sprint-2m"
+ }
+}

policy.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acb1bcb8ed11b5c93ebb89df2ed9ae6b2c440016b09e625964cef700a2242990
+size 831318