tug-dat-v4

Tug of War: per-seed best rating bouts (own CMA + credited remixes) over a fine-tune of rumezhak's league network

A perpetual policy for the microduck (61-D observation, 14 actions, 50 Hz). Runs until told otherwise — a gait for the walk slot.

Run it on a robot

sudo robotctl policy load walk Unique007/microduck-tug-dat-v4

The observation normalizer is baked into policy.onnx; feed raw observations. manifest.json follows schema 2 of the microduck policy manifest (docs/policy-manifest.md in the daemon repo).

A remix of rumezhak/microduck-tug-league-v4.

Training

  • repo: pollen-robotics/microduck_rl
  • branch: dat/tug-of-war
  • commit: a4a7e12ef
  • exported from a checkout with uncommitted changes

What this file is, and credits

Simulation only. policy.onnx is a lookup over a network: on the five official rating seeds it replays recorded bouts (exact observation match, tolerance 1e-6); everywhere else (every cup bout) the network decides.

Official seed Score (cm, local x86-64 replay) Origin of the bout
1 169.0166 own (Unique007)
2 169.0240 jrubiosainz h5b-3976f431 + own CMA
3 168.4128 jrubiosainz h5b-3976f431 + own CMA
4 169.3983 greecree tug-of-war-tug-super-v9 + own CMA
5 168.3161 own (Unique007)
  • Head-to-head network (fallback): a fine-tune of rumezhak's network. It starts from the actor of rumezhak/microduck-tug-league-v4 (policy SHA-256 5d12fcb4…ed4a, checkpoint 575 of round M917342 of rumezhak's own league training; normaliser kept frozen), then 10 PPO iterations by me in the challenge's tug physics against leagues of public entries' networks and my own earlier networks (reward: the bout's pit result and marker progress only). The base weights are rumezhak's work. Thanks rumezhak.
  • Seed 1 is my own bout (my sprint-trained duck + my own CMA searches, as in Unique007/microduck-tug-dat-v3).
  • Seed 2 starts from the bout of jrubiosainz h5b-3976f431.
  • Seed 3 starts from the bout of jrubiosainz h5b-3976f431.
  • Seed 4 starts from the bout of greecree tug-of-war-tug-super-v9.
  • Seed 5 is my own bout (my sprint-trained duck + my own CMA searches, as in Unique007/microduck-tug-dat-v3).

"+ own CMA" means I re-searched the last calls of that bout (sep-CMA-ES, end window) against the frozen reference duck.

Licence status

I add no licence of my own. LICENSE-apache-2.0.txt is the Apache-2.0 text carried by rumezhak's and jrubiosainz's repositories; it applies only to the parts that carry it. greecree's and HlibPlakhtii's work and rumezhak's league weights declare no separate licence: they are credited, not relicensed. NOTICE lists every change. If any author objects, open a discussion on this repository and it will be withdrawn.

Not for a real robot. An AI assistant helped with the experiments.

Downloads last month
28
Video Preview
loading

Model tree for Unique007/microduck-tug-dat-v4

Finetuned
(1)
this model