hmdliu/fern-juniper-falcon

easy-wam-exp: MiniWAM re-implementation on EasyWAM

An unofficial re-implementation of MiniWAM (paper, project page; no official code released) built on top of EasyWAM (pinned as a submodule).

MiniWAM trains a small world-action model in two stages:

1. Stage 1 (PRISM). An encoder compresses current + future visual features (Wan2.1 VAE latents) into a compact latent grid (64 x 3 x 4 x 4 per camera). It is trained with an inverse-dynamics action flow loss, a feature reconstruction loss and a small KL term.
2. Stage 2. A freshly initialised two-expert (world / action) transformer jointly denoises the frozen PRISM latents and a 16-step action chunk. It is conditioned on the current frame, the language instruction and the proprioception.

Setup and usage: see docs/SETUP.md.

Results so far (LIBERO, Wan2.1-VAE features)

Success rate (%) over 10 tasks x 50 trials per suite. Single training seed, last checkpoint.

Model (training data)	Spatial	Object	Goal	Long	Avg
MiniWAM, Wan + PRISM (all 4 suites)	97.8	99.2	94.6	90.2	95.5
MiniWAM, Wan + PRISM (Object only)	–	99.6	–	–	–
Control: noise world tokens, no world loss (Object only)	–	94.6	–	–	–
MiniWAM, Wan + PRISM (Long only)	–	–	–	91.2	–
Paper, Wan + PRISM (all 4 suites)					94.8

Full-data training cost: Stage 1 ~1.1 h + Stage 2 ~6.4 h on 2x H100/H200, plus ~0.5 h of parallel eval. Details are in docs/records/2026-10-19_miniwam_libero_wave1.md (Chinese).

Repository layout

miniwam/                 MiniWAM code (model, PRISM, data, feature extraction, eval)
  model.py               MiniWAM (Stage 1 / Stage 2 / noise-control) + flow-matching losses + inference
  prism.py               PRISM encoder / decoder
  world_dit.py           world expert (current-frame + latent tokens, 3D RoPE)
  data.py                dataset over cached features (+ LeRobot actions / proprio / text)
  extract_wan_features.py  offline Wan2.1-VAE window features
  export_targets.py      freeze Stage 1, cache posterior means as Stage-2 targets
  eval_libero.py         LIBERO closed-loop eval (EasyWAM evaluator, trial-chunk sharding)
  summarize_eval.py      aggregate eval results
  tests/                 sanity checks (features / model / data)
configs/                 Hydra configs (train.yaml, eval_libero.yaml, data/, model/, task/)
scripts/setup/           environment setup + downloads
scripts/miniwam/         job scripts (run with sbatch or plain bash); train_chain.sbatch = S1+export+S2
                         in one 2-GPU job; launch_main.sh = train_chain + 1-GPU eval jobs
scripts/easywam_smoke/   legacy: smoke test of the released EasyWAM checkpoint (not needed)
docs/SETUP.md            setup / reproduction / extension guide
docs/report/             paper-derived implementation spec (Chinese)
docs/plans/              reproduction plan + spec-vs-code cross-check (Chinese)
docs/records/            experiment records (Chinese)
third_party/EasyWAM      submodule @ a778b6f
third_party/LIBERO       submodule @ 8f1084e