---
license: apache-2.0
base_model: allenai/MolmoAct2-BimanualYAM
tags:
- lerobot
- molmoact2
- robotics
- bimanual
- yam
pipeline_tag: robotics
---

# MolmoAct2 — Bimanual YAM cube-stacking (action-expert, chunk_size=30)

Fine-tune of [`allenai/MolmoAct2-BimanualYAM`](https://huggingface.co/allenai/MolmoAct2-BimanualYAM)
on the `atharva-pantheon/yam-stack-cube` dataset (44 bimanual-YAM demos, 3 cameras, 14-dim
absolute joint pose, 10 fps). Task instruction: **"stack the cubes"**.

Trained with LeRobot (`molmoact2-policy` branch) for a **smooth, RL-ready** continuous policy.

## Recipe
- Action-expert-only, continuous (VLM frozen; ~578M trainable / 5.4B total). No LoRA on the action expert.
- `chunk_size=30`, `n_action_steps=30` (3 s lookahead @ 10 fps → long, smooth motion).
- `setup_type="bimanual yam robotic arms in molmoact2"`, `control_mode="absolute joint pose"`.
- bf16, 8 flow timesteps, action-expert lr 5e-5, cosine schedule (200 warmup), batch_size 8.
- Normalization **reused from the base checkpoint's `yam_dual_molmoact2` tag** (not recomputed on
  the 44-demo set, whose joint range is much narrower) for scale-consistent, smooth actions.

## Checkpoints
- `20k_run/checkpoint_020000/` — 20,000-step run, final loss ≈ 0.009.
- `40k_run/checkpoint_005000` … `checkpoint_040000/` — 40,000-step run, every 5k steps.

Each folder is a LeRobot `pretrained_model` (weights + pre/post processors). Pick the best
checkpoint on the physical robot (no simulator for YAM). Evaluation is on hardware.

## Training loss (20k run)

![20k training loss](assets/loss_curve_20k.png)

Flow-matching loss decays smoothly 0.093 → 0.009 (log scale), flattening after ~15k steps;
no instability/spikes. The 40k-run curve is added under `assets/` when that run completes.