alam_plus_pi_libero

ALAM + pi0 post-trained for LIBERO

Repository ID: Mark-ZJTang/alam_plus_pi_libero.

Inference-only step-30,000 Orbax checkpoint used as the shared checkpoint for the LIBERO Table 9 release configuration. Training effective horizon is 20. At inference, Spatial/Object use H=14 and Goal/Long use H=18; replan steps are 5, 10, 7, and 12 respectively.

The public full post-training resource contract is 8 GPUs. Inference and the documented CUDA/EGL acceptance use one sufficiently large idle GPU.

The checkpoint includes params, _CHECKPOINT_METADATA, and assets/libero_real/norm_stats.json; optimizer train_state is intentionally excluded. Both LIBERO and MetaWorld use the alam_pretrain_latent_action_tokenizer, alam_pretrain_latent_action_tokenizer. The project owner confirms that this shared tokenizer configuration has passed LIBERO validation.

The available historical evidence supports the shared-checkpoint release configuration, but the Spatial server identity was not recorded in its client log; this provenance limitation is documented in the GitHub release.

From the matching GitHub code checkout:

.venvs/publish/bin/python workflows/publishing/download_huggingface.py \
  --artifact alam_pretrain \
  --artifact alam_plus_pi_libero
bash workflows/libero_evaluation/evaluate_all_suites.sh --gpu 0

Each suite uses 10 tasks x 50 trials. Recorded successes are Spatial 496/500 (99.2%), Object 498/500 (99.6%), Goal 495/500 (99.0%), and Long 472/500 (94.4%), for a 98.05% mean reported as 98.1%.

The scores above are the recorded historical benchmark results. The shared tokenizer validation status is confirmed by the project owner; no new numerical benchmark results are introduced in this model-card update.

For existing code checkouts, set both ALAM_METAWORLD_TOKENIZER and ALAM_LIBERO_TOKENIZER in .env to evaluation/checkpoints/alam/alam_pretrain_latent_action_tokenizer.