npow/microduck-coop-mappo

Microduck cooperative-pushing MAPPO checkpoints

This repository preserves training checkpoints for the Cargo Ducks / Microduck cooperative-pushing research workbench. Training code and the experiment record are in npow/microduck-coop.

Important status

The V97 continuation was paused by the user to free the GPU. V97 is unfinished and has not been evaluated. Eight of twelve branches have checkpoints at update 23,180; four branches had not started. These are resumable training states, not final policies or evidence of a training-horizon benefit. Do not cite V97 as a result. Its matched 60-second and 120-second actor weights differed before pause, confirming that the corrected environment cap affects training; task delivery still requires the planned held-out evaluation.

The earlier V96 horizon comparison is invalid because an environment reset overwrote the requested cap. V96 artifacts are intentionally not included.

Files

• v82-source-checkpoints/: six peer-visible and trained-no-peer source checkpoints at update 23,100 (three source seeds each). They are the initialization points for V97 and are useful baselines, but are not claimed to solve the task.
• v97-partial-checkpoints/: eight V97 checkpoints at update 23,180, with optimizer and RNG state for continuation. Four planned branches are absent because they had not started.
• v97-training-logs/ and v97-episode-metrics/: progress logs and training episode metrics for the eight partial branches. These are training diagnostics, not held-out success estimates.
• v97-status.json: pause state.
• manifest.json: file sizes and SHA-256 checksums, plus checkpoint metadata.

Resuming

Use the source checkpoint and V97 runner from the linked GitHub repository. The runner needs to be updated to resume the eight partial checkpoints at 23,180, train the four missing branches, and then run all 24 held-out evaluations before producing the paired analysis. The current published runner describes the campaign but does not yet automate resuming this paused state. All six V82 sources are included so the original machine paths are not required.

The MAPPO high-level checkpoint depends on the frozen locomotion policies and robot model assets in the GitHub repository; it is not a standalone robot controller.

Loading

These files are PyTorch torch.save checkpoint dictionaries containing model parameters, optimizer state, random-number-generator state, and training arguments. They are intended for continuation with the matching training code, not direct use as a generic Hugging Face Transformers model. Only load PyTorch pickle checkpoints from sources you trust.

License: Apache-2.0.