Describe the selected release candidate, usage, and validation limits

README.md

@@ -0,0 +1,111 @@
+---
+language: en
+tags:
+  - robotics
+  - libero
+  - video-to-action
+  - openwam
+pipeline_tag: robotics
+---
+
+# OpenWAM VTA LIBERO-10
+
+A dual-expert video-then-action policy for LIBERO-10. This release contains
+the complete video and action experts from the selected step-15,000 checkpoint.
+It is not an inverse-dynamics-only model or a video-backbone-only export.
+
+The policy first predicts future video latents from the task prompt and observed
+scene, then predicts robot actions conditioned on those latents. During
+closed-loop execution, new real observations update its recurrent history.
+This is a single trained VTA policy, not a composition with an external IDM.
+
+**Status:** private release candidate at
+[`OpenWAM-Stanford/OpenWAM-VTA-LIBERO10`](https://huggingface.co/OpenWAM-Stanford/OpenWAM-VTA-LIBERO10).
+The selected checkpoint is step **15,000**, replacing the earlier proposed
+VTA release. The inference weights contain approximately 7.67 billion FP32
+parameter elements (about 30.7 GB); this is not an inference-memory estimate.
+
+Project: https://openwam.github.io/
+
+Code: https://github.com/OpenWAM/OpenWAM
+
+## Inference Contract
+
+- Program: `video_then_action`, with `video_only` recurrent history visibility.
+- Two 128x128 RGB camera views, concatenated into a 128x256 frontend input.
+- Four predicted latent frames per chunk and sixteen raw 7D action commands.
+- Attention window: 30; 20 video and 20 action denoising steps.
+- Video guidance scale: 5; conditioned action generation with scale 1.
+- Per-chunk proprioception, single-frame condition latents, and the
+  `legacy_prefix_single_frame_perchunk_proprio` sequence contract.
+
+Use `resolved_config.yaml` with `model_state.pt` through the maintained
+OpenWAM checkpoint loading path. Configure the external LingBot-VA VAE,
+text encoder, and tokenizer assets using `paths.models.lingbot_va_base`.
+Their source revision is recorded in `release_metadata.json`. The full
+checkpoint supplies the policy weights; no private initialization checkpoint
+is needed for inference.
+
+## Download And Run
+
+Authenticate with an account that can read this private repository, then download
+the inference bundle:
+
+```bash
+hf download OpenWAM-Stanford/OpenWAM-VTA-LIBERO10 --local-dir vta-libero10
+cd vta-libero10
+sha256sum --check SHA256SUMS
+```
+
+Install OpenWAM and its LIBERO simulator dependencies following the code
+repository instructions. Use the code revision recorded in
+`release_metadata.json`, and set `paths.models.lingbot_va_base` in
+`configs/local_paths.yaml` to the matching frontend assets. Run from that
+checkout with `CKPT` set to the downloaded bundle's absolute path:
+
+```bash
+MUJOCO_GL=osmesa python scripts/run_libero_policy.py \
+  --cfg "$CKPT/resolved_config.yaml" --checkpoint "$CKPT/model_state.pt" \
+  --benchmark libero_10 --task-id 0 --episode-idx 0 --seed 0 \
+  --frontend-encode-mode lingbot_streaming_vae \
+  --inference-window-size 30 --execute-frame-chunk-size 4 \
+  --startup-model-obs-frames 1 --startup-env-init-steps 5 \
+  --max-timestep 800 --max-chunks 50 \
+  --runtime-device cuda:0 --action-device cuda:0 \
+  --frontend-device cuda:0 --decode-device cuda:0 \
+  --save-rollout-video --output-dir outputs/vta-libero10
+```
+
+This example uses the maintained CLI and preserves the checkpoint's inference
+contract. A fresh GPU rollout of this packaged checkpoint remains to be run;
+the example is not a new benchmark result. Do not use partial-checkpoint loading
+or substitute another checkpoint's inference configuration.
+
+## Contents And Evidence
+
+| File | Contents |
+| --- | --- |
+| `model_state.pt` | Tensor-only policy checkpoint, including both experts |
+| `resolved_config.yaml` | Portable architecture and inference settings |
+| `release_metadata.json` | Code/assets revisions, tensor counts, source/export hashes |
+| `SHA256SUMS` | Integrity checks for all release files |
+| `README.md` | This model card |
+
+The bundle contains tensor-only policy weights, a portable inference config,
+checksums, and release metadata. It does not contain optimizer state, training
+logs, or private dataset paths. Every active policy tensor retains its exact
+original values and dtype. The export omits only the two tensors belonging to
+the retired post-decoded feature projection (`visual_tower.decoder.proj`),
+which is absent from the maintained VTA inference architecture. Neither the
+VAE nor either expert is removed. Source and exported file hashes are recorded
+separately; the exported file is not byte-identical to the training archive.
+
+No benchmark success rate is assigned to this checkpoint by this release.
+Results measured with other VTA checkpoints or other inference overrides must
+not be attributed to these weights. Architecture/configuration checks establish
+loadability, not closed-loop policy performance.
+
+Weight-license confirmation and a rollout smoke test remain release gates.
+The policy was developed for simulated LIBERO manipulation; no real-robot
+safety or generalization guarantee is made. Private availability does not
+constitute a redistribution license or approval for deployment.