lewm-rs β€” SO-100 Trained Checkpoint

A Rust/Burn implementation of LeWorldModel (Le-WM) trained on the SO-100 pick-and-place dataset.

This checkpoint is trained on abdelstark/so100-pickplace-lewm-ready (1.9 GB HDF5, 6,559 timesteps, 50 episodes at 10 fps) for 6-DOF robotic manipulation.

Training Results

Metric Value
Steps 5,000
Wall time 864 s (~14 min) on A10G-large
Initial loss 0.500
Final loss 9.56e-05
Gradient explosions 0
Job abdelstark/6a070e02e48bea4538b9e2a5 (v11a)

Loss curve:

Step Total loss SIGReg Pred loss
1 5.00e-01 5.00e-01 2.40e-04
1,000 2.03e-01 2.03e-01 8.05e-06
2,500 3.70e-04 3.69e-04 1.44e-06
5,000 9.56e-05 9.50e-05 5.34e-07

Architecture

Same ViT-Tiny JEPA architecture as the PushT model, adapted for 6-DOF robotic manipulation:

  • ViT-Tiny visual encoder β€” 192-dim, 12 transformer layers, 3 attention heads, 14Γ—14 patch tokens from 224Γ—224 top-view images
  • Action encoder β€” 6-DOF joint actions (smoothed to 10-dim) β†’ 192-dim embeddings
  • AdaLN-zero autoregressive predictor β€” 6 transformer blocks, 16 heads, 2048 MLP
  • Projector / Pred-proj MLPs β€” 192 β†’ 2048 β†’ 192 with BatchNorm1d

Total parameters: ~18M

Encoder:   ViT-Tiny (192-d, 12L, 3H, patch=14, img=224, top-view camera)
Predictor: 6 blocks, 16 heads, dim=192, mlp=2048
Action:    6-DOF β†’ smooth10 β†’ emb192 (mlp_scale=4)
Training:  SIGReg + prediction MSE loss (Ξ»=1.0, knots=17, proj=1024)

Training Details

Field Value
Dataset abdelstark/so100-pickplace-lewm-ready
Hardware A10G-large (HuggingFace Jobs)
Precision bf16 mixed
Steps 5,000
Batch size 64
Optimizer AdamW (lr=3e-4 β†’ 1e-5, warmup=500, wd=0.05, Ξ²=[0.9, 0.95])
Grad clip 1.0
History size 3 frames
Prediction horizon 4 frames
Camera Top view (224Γ—224)
Seed 0

Dataset

Property Value
Source abdelstark/so100-pickplace-lewm-ready
Episodes 50
Timesteps 6,559
Sampling rate 10 fps
Image resolution 224Γ—224
Action space 6-DOF (joint angles)
HDF5 size 1.9 GB

Artifacts

File Description
train/so100-full-20260515T122820Z/step_0005000.safetensors Model weights
train/so100-full-20260515T122820Z/step_0005000.mpk Full checkpoint
train/so100-full-20260515T122820Z/train_report.json Training summary
train/so100-full-20260515T122820Z/train_losses.jsonl Per-step loss log

Warm-Start Evaluation

Pending. The .mpk checkpoint is available for warm-start evaluation (initialising from this SO-100 checkpoint vs. from the PushT checkpoint).

Repository

License

MIT. See LICENSE.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading