CIS6280 World Models for Stacking: checkpoints (temporary backup)
Checkpoints, training logs, and evaluation reports for a MuJoCo cube-stacking
benchmark (dexterous hand, two 128 px camera views). Datasets live in
YongYong/cis6280-stack-datasets.
INDEX.md is the full experiment log.
All .pt files are torch.save dictionaries (state_dict, config, info) and need
torch.load(..., weights_only=False). The model code (cube_pick/stack_*.py) is in the
(private) world-model-gym repository, branch stack-world-model-baseline.
Contents
| Folder | What |
|---|---|
bc/r0, bc/r1 |
Privileged-state BC teacher (MLP on 190-dim features, action chunks). Data collection used bc/r1/F2-c10-iid-100k/step-050000.pt (2-cube success 94% on val). |
jepa/ |
JEPA world models on tiled front|side frames (128x256, ViT-Tiny patch 8): lewm-c3, lewm-c80 (LeWM, SIGReg), smwm-c3, smwm-c80 (SMWM, inverse dynamics), lewm-p-c3 (LeWM with a proprio token). model.pt = final. |
policy/ |
Diffusion-forcing latent policies on frozen SMWM-c3 latents (16 action blocks of 5 raw actions): df-smwm-c1, df-smwm-c2 (all episodes), df-smwm-c1-sigma04 (episodes with action noise sigma <= 0.04). policy.pt = final EMA weights; eval/ = closed-loop reports and videos. |
tdmpc/ |
Offline TD-MPC2: tdmpc-rgb, tdmpc-latent (reference objective), tdmpc-latent-sarsa (dataset next action in the TD target). aborted/ = partial runs incl. TD3+BC variants. |
ppo/ |
PPO runs (checkpoints and logs; training rollouts not included). |
wm-baselines/ |
Earlier world-model baselines. |
goal-planning/ |
Goal-conditioned selection among DF policy samples with SMWM rollouts (stage goals vs dense subgoals). |
Headline results (val scenes, closed loop)
- DF policy, 2-cube task: 16-18% success with all episodes; 43% when trained on sigma <= 0.04 episodes.
- Offline TD-MPC2: 0% in all variants; the reference TD target overestimates Q by ~100x on this data.
- Goal-conditioned planning (4-cube task, DF C=1): dense subgoals every 8 s raise "at least one cube" from 58% (no planning) to 67-75%.