3DWM world model — FrankaPickPlace3D (epoch 28)
The trained 3D point-cloud world model used for closed-loop MPPI planning on
kinder's MuJoCo FrankaPickPlace3D-o1 task (reach, grasp, carry, place). Every
number in the accompanying reproduction package was measured on this checkpoint.
The model predicts the next per-part point cloud; MPPI scores sampled action sequences by rolling it forward. No policy is trained — the planner is the controller.
File
franka_pickplace_rollout10_bs16_dt01_epoch28_model.pt — 200 MB,
md5 01ec30be1f1c4ff94a43e83f1591b413, 16.65 M parameters.
Contents: {'epoch': 28, 'dynamics': <state dict>, 'optim_state_dict': ...}, so
training can be resumed from it as well.
Training
| dataset | Franka-pickplace-1000demos-v2 (1000 demos, 109,084 frames) |
| architecture | PTv3 encoder/decoder, multi_frame_num=2, 7 action channels |
| rollout length | 10 (autoregressive) — the planner's horizon must match |
| dt | 0.01 |
| batch size | 16 |
| optimizer | AdamW, lr 1e-4, StepLR (milestones 10/20) |
| total epochs | 60 — epoch 28 is a mid-training checkpoint; later epochs were not better for planning |
Use
Place it where the plan config resolves checkpoints (tools/utils.py appends the
dataset scenario to exp_name):
logdir/3dwm_franka_local/franka_pickplace_rollout10_1000demos_v2_bs16_dt01_pickplace/epoch_28/model.pt
Then follow section 4 of the reproduction guide in the code repository.
Two things to know
The MuJoCo environment must be pinned. The demos were generated against
kindergarden commit
1d0cc56,
and no published version of that package reproduces it. PyPI
kindergarden==0.2.0 carries the same version string but was built three days
earlier, before the FR3 pinch_site the demos' actions are defined on existed.
v0.2.1 and later moved the FR3 mount (mount_height → mount_surface), which
changes every scene and puts this model out of distribution. Both fail silently.
The code repository ships scripts/check_kinder_env.py to verify by content hash.
Runs are not bitwise reproducible even with this checkpoint and a fixed seed — GPU nondeterminism moves per-episode event steps. Reproduce rates and rankings over a seed range, not individual step numbers. On 15 seeds this checkpoint places the cube 15/15.