alam_plus_pi_metaworld_mt50

ALAM + pi0 post-trained for MetaWorld MT50

Repository ID: Mark-ZJTang/alam_plus_pi_metaworld_mt50.

Inference-only step-30,000 Orbax checkpoint for the released MetaWorld MT50 evaluation. The current release configuration uses raw action horizon 6, effective inference horizon 5, replan step 5, and camera position (0.71, 0.075, 0.70).

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

The checkpoint includes params, _CHECKPOINT_METADATA, and assets/metaworld_mt50/norm_stats.json; optimizer train_state is intentionally excluded. Both MetaWorld and LIBERO use the alam_pretrain_latent_action_tokenizer, alam_pretrain_latent_action_tokenizer.

See the GitHub release profile for the current evaluation configuration and its recorded results.

From the matching GitHub code checkout:

.venvs/publish/bin/python workflows/publishing/download_huggingface.py \
  --artifact alam_pretrain \
  --artifact alam_plus_pi_metaworld_mt50
bash workflows/metaworld_evaluation/evaluate_mt50.sh --gpu 0

The frozen 50-task protocol uses 10 episodes per task. The historical camera-x=0.70 result is 434/500 episode successes (86.8% episode-weighted) and 84.98% after macro averaging Easy/Medium/Hard/Very-Hard groups, reported as 85.0%.

The recorded current camera-x=0.71 run achieved 432/500 (86.4% episode-weighted) and 85.366883% difficulty-macro success. These are separate recorded runs.

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