A²-MEAS
Inference resources for A²-WLA and A²-MEAS, from Toward the Era of Agentic Embodied Intelligence.
Use these resources with the matching A2MEAS evaluation source release.
| Path | Contents |
|---|---|
weights/ |
One compressed BEHAVIOR inference store for 25 policies |
behavior/policy/ |
Distinct BEHAVIOR policy configs, processors and normalization statistics |
behavior/routing.json |
Routing for 100 tasks, with N=2 |
robotwin/policy/ |
RoboTwin inference checkpoint, config and processors |
adapter/ |
BEHAVIOR Adapter weights and twelve-task configuration |
assets/rynnbrain/ |
Shared tokenizer and image processing files |
resource-manifest.json |
Published file sizes, SHA-256 values and restored checkpoint identities |
The BEHAVIOR store contains only inference tensors and the reference tensors
needed to reconstruct them. The evaluation downloader restores standard
safetensors checkpoints and verifies the restored SHA-256 values. RoboTwin
loads its safetensors checkpoint directly. No quantization or numeric changes
are applied by this packaging step.
From the A2MEAS checkout, after installing its environments and authenticating for any gated third-party resources:
runtime/model/bin/python serve.py behavior prepare --adapter
runtime/model/bin/python serve.py robotwin prepare --sam
Both commands use pinned revisions and populate the checkout's resources/
directory. The source setup.sh prepares environments, SDKs and simulator assets.
SAM and the Adapter judge are downloaded separately from their official model
repositories. BEHAVIOR datasets and their decryption key must be obtained through
the official authorized download flow.
A²-WLA inference weights are licensed under Apache-2.0; see LICENSE and NOTICE. Applicable upstream terms remain in force for third-party components.
The embedded DINOv3 vision encoder is distributed under its DINOv3 License.