TaskAnchor–π0.5: RoboMemArena 50k weights

This repository contains the complete inference checkpoint and its required configuration, tokenizer, and normalization assets. Source code will be published separately on GitHub; the link will be added here when available. No benchmark scores are published in this repository.

Files

50000/
  params/                           # Complete backbone + TaskAnchor, Orbax OCDBT
  assets/robomem_memory.json
  assets/pi05_libero/norm_stats.json
  _CHECKPOINT_METADATA
recipe.json                         # Keep one level above 50000/
tokenizer/paligemma_tokenizer.model
weights_manifest.json               # File hashes and tensor inventory
SHA256SUMS                          # Inference payload checksums
release_manifest.json               # Full publication inventory
LICENSE_GEMMA.txt
NOTICE

The inference payload is approximately 10.86 GiB. It contains full model weights, not an adapter-only delta. The parameters are stored in native JAX/Orbax OCDBT format; they are not Safetensors. The checkpoint was restored on CPU and checked for finite values. Optimizer state is not included.

Download

hf download henry13522/Taskanchor --local-dir ./Taskanchor
cd Taskanchor
sha256sum -c SHA256SUMS

In the matching TaskAnchor inference code, set the checkpoint path to /absolute/path/to/Taskanchor/50000 and retain recipe.json in its parent. Set TASKANCHOR_TOKENIZER_PATH to the absolute path of the supplied tokenizer/paligemma_tokenizer.model. Normalization is read from the checkpoint assets. The full checkpoint does not require an additional base-weight download for inference.

Configuration

The main view uses refined = current + offset, with a 1152→512→1152 projection, eight temporal layers, eight heads, and dropout 0.10. The wrist view remains a current observation. Images are 224×224. Historical visual features are pooled to 4×4 tokens per frame. This checkpoint retains a separate permanent frame0 plus up to 191 history frames, with history stride 10.

Inference predicts 20 actions and executes 10 before replanning. Continuous progress predictions are mapped to the nearest task-specific stage coordinate before conditioning, with two decimal places. The 50k snapshot corresponds to 5k adapter-only steps followed by 45k joint steps; the visual encoder remains frozen. The recorded global training batch is 256. Full numerical settings are retained in recipe.json and the saved checkpoint metadata.

License

The full pretrained-model checkpoint retains applicable upstream π0.5 / PaliGemma / Gemma terms. See LICENSE_GEMMA.txt and NOTICE. The separately published source code uses Apache-2.0; that source license does not replace the pretrained-weight terms.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

Upstream: OpenPI and RoboMemArena.