ActMem-VLA — LIBERO-Mem representative plugins

Ten corrected task-specific plugins, one for each LIBERO-Mem task. Every representative uses training seed 42, checkpoint 1500, memory width/depth 1024/2, 4-layer PreAction Expert (PAE) and 4 PAE + 6 frozen AE denoising calls. They are a uniform reference set, not a claim of task-wise optimal checkpoint selection.

Important: these are plugin weights

plugin.safetensors does not include the frozen VLM or original Action Expert. Use the matching jointly fine-tuned pi0.5 LIBERO-Mem base and its normalization assets. An arbitrary pretrained pi0.5 checkpoint is not an equivalent substitute. The exact base identity is preserved in each metadata.json. The metadata's base_checkpoint_checksum and the SHA-256 of the converted base model.safetensors are different identifiers and must not be confused.

The original base URI is a server-local provenance path, not a public download link. Unless a base/ directory is present, this repository is a plugin release and is not standalone inference-ready. No third-party weight license is overridden here.

Layout

checkpoints/task_01/seed_42/step_1500/
    plugin.safetensors
    metadata.json
    SHA256SUMS
... task_02 through task_10 ...
representative_checkpoints.json
release_manifest.json
code/actmem_release_20260928.tar.gz

release_manifest.json records remote-verified SHA-256 values, byte counts and the publication revision. Always check it to confirm upload completion.

Code and experimental evidence

Project: https://github.com/hit-my/PhaseVLA (method name: ActMem-VLA). The code/ archive preserves the complete published code/data release at commit 68a7610daba626c893dbb891587d4f0f9cb99129. GitHub publication is verified: https://github.com/hit-my/PhaseVLA/commit/68a7610daba626c893dbb891587d4f0f9cb99129

Inside the archive, releases/2026-09-28/ contains corrected runtime wrappers, immutable provenance snapshots, audited episode-level records from A100/LGD, the separate paper-analysis snapshot, and plotting inputs. Training/evaluation wrappers preserve historical absolute paths and require adaptation to your environment. No new training or simulator run was performed for this publication.

Loading

1. Download the repository or just one task folder using huggingface_hub.snapshot_download.
2. Install the project environment specified by the code archive; configure the matching fine-tuned base and normalization assets.
3. Use the project's FutureMamba checkpoint/policy loading path, with the task-specific plugin directory and its metadata. Keep action normalization, execution horizon, denoising order and recurrent reset behavior unchanged.
4. Reset recurrent history at episode start. Append only executed commands, not the unused suffix of predicted actions.

Example download (the downloaded plugin alone cannot run a policy):

from huggingface_hub import snapshot_download
path = snapshot_download(
    repo_id='HITdongdong/ActMem-VLA',
    allow_patterns=['checkpoints/task_06/seed_42/step_1500/*',
                    'representative_checkpoints.json', 'release_manifest.json'],
)

Evaluation and limitations

The experiment archive separates task, training seed, checkpoint, evaluation seed and per-episode success/steps. Evaluate against the same initial-state protocol and episode budget; three evaluation seeds are not three independently trained policies. Historical and repaired memory-gradient runs must not be merged. Legacy fine-tuned-AE no-PE and later frozen-AE no-PAE are different comparisons.

No single aggregate score is advertised for this representative subset. Paper figures may use other checkpoint-selection rules; consult their source manifests. Existing memory visualizations are offline demonstration replays, not proof of long-term causal memory. Physical robot deployment needs task-specific validation and supervision.