so101-stackcube-dreamer-seed0

sim_vla dreamer arm (Dreamer world model + SmolVLA, no graph), trained on SO-101 real data, task stackcube (blue cube on red, episodes 50-98 of hungho77/so101-multitask).

• Code: chickbong221/r2dreamer-graph, branch so101-real, commit 920d07dbfedb6d19e34c19736d93ee2b5abf66d0
• Hardware: 1x H100 MIG 3g.40gb. batch size 8 (the H100 config uses 16), so each stage sees half the samples of the reference run.
• Stage 1A world model: 15000 steps; stage 1B imitation: 15000 steps (final loss ~0.03). No online stage.

python -m sim_vla.training.pipeline \
  --data real --task stackcube --experiment dreamer --batch-size 8 \
  --world-steps 15000 --imitation-steps 15000 --online-steps 0 \
  --world-lr 1e-4 --world-warmup-steps 1000 --world-final-lr 1e-5 \
  --imitation-lr 1e-4 --imitation-warmup-steps 1000 --imitation-final-lr 2.5e-6 \
  --seed 0 --device cuda \
  --save-checkpoints --out logdir/sim_vla/real/stackcube/dreamer_seed0

Files: world_model.pt, imitation.pt (+ their .json metadata), normalization.json, train.log. Load with --resume-from <this folder>.