Bi-LaWM RoboTwin Stage3-Only — Policy From Scratch
Repository: GT-111/bi-lawm-robotwin-stage3-only (renamed from lawam-robotwin-steps68000 on 2026-08-31; the old URL redirects here).
Intermediate checkpoints from the RoboTwin Stage3-only policy-from-scratch training run of Bi-LaWM.
The Stage3 policy was trained without a pretrained policy checkpoint. It uses the RoboTwin Stage1 BiLAM export as its latent-action model initialization, so “from scratch” here refers specifically to the policy training stage rather than the complete Bi-LaWM stack.
Checkpoints
| Hub path | Training step | Format |
|---|---|---|
checkpoints/step-010000/ |
10,000 | Full sharded FSDP training checkpoint |
checkpoints/step-020000/ |
20,000 | Full sharded FSDP training checkpoint |
checkpoints/step-030000/ |
30,000 | Full sharded FSDP training checkpoint |
checkpoints/step-040000/ |
40,000 | Full sharded FSDP training checkpoint |
checkpoints/step-050000/ |
50,000 | Full sharded FSDP training checkpoint |
checkpoints/step-060000/ |
60,000 | Full sharded FSDP training checkpoint |
checkpoints/step-068000/ |
68,000 | Latest published full sharded FSDP training checkpoint |
Each checkpoint contains the model shards, per-rank optimizer and RNG states, scheduler state, and trainer_state.json. These are resumable distributed-training checkpoints rather than merged inference-only weights.
Experiment
- Project: Bi-LaWM
- Dataset/environment: RoboTwin
- Stage: Stage3 policy training
- Policy initialization: from scratch (no pretrained policy checkpoint)
- Latent-action initialization: RoboTwin Stage1 BiLAM export
- Training plan: 100k steps
- Published interval: every 10k steps, plus the latest available 68k checkpoint
Loading
Use these checkpoints with the matching Bi-LaWM code and FSDP configuration. For inference or evaluation, export/merge the selected distributed checkpoint with the project checkpoint tooling first.