towel_fold_smolvla_shaping_005000

1goldexperience1/towel_fold_smolvla_shaping_005000

towel_fold_smolvla_shaping_005000 is a reinforcement learning model from 1goldexperience1. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for lerobot. The card lists the license as apache-2.0.

SmolVLA (SmolVLM2-500M vision-language backbone + action-expert head) policy trained to shape a real dual-Piper towel-folding skill. It is not a standalone controller by design: it runs as the SmolVLA candidate under…

Downloads · 30 days

29% of all-time downloads

All-time downloads

Public

Parameters

450M

907 MB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors907 MB · 100%

Parameter types

How the weights are stored.

BF16447M · 99%

Model card

From the Hugging Face model README

SmolVLA towel-folding shaping policy (HQ60, 5,000 steps)

SmolVLA (SmolVLM2-500M vision-language backbone + action-expert head) policy trained to shape a real dual-Piper towel-folding skill. It is not a standalone controller by design: it runs as the SmolVLA candidate under a continuous authority blend on top of the proven ACT safety base (see piper-dual-arm-smolvla-towel-folding).

Training


Type	LeRobot `smolvla` policy (SmolVLM2-500M + action expert)
Fine-tune	5,000 steps, batch size 8, from `smolvla_towel120_continue20k_b8_50k_v4` @ step 050000
Recipe	SmolVLA Stage B ("one-more-step"): vision encoder frozen, action expert only
Data	`1goldexperience1/towel_fold_dataset_aug_v1` — 120 real dual-arm demos, 85,187 frames @ 30 FPS (HQ60 new-only subset)
Action chunk	50 (n_action_steps=50, n_obs_steps=1)
Seed	1000

Observation / action spec

Input — 3 RGB views camera1/2/3 resized to 256×256 + 28-dim dual-arm joint state (2 arms × 7 motors × position+effort interleaved).
Output — 14-dim absolute joint-position targets (left j1..j6 + left gripper, right j1..j6 + right gripper).

How it is used on the real robot

This checkpoint is the SmolVLA candidate loaded by scripts/run_hybrid_towel_blend750.sh on the rig (SMOL_MODEL), blended every 30 Hz tick with the ACT base released as 1goldexperience1/towel_fold_act_v4_040000:

target_arm = safe_base_arm + authority * (smol_ema_arm - safe_base_arm)

authority rises to the cap (0.9) only while both policies agree; decays on soft/hard reject.
Corrections are magnitude-capped (<=0.15 rad) and slew-limited (<=0.02 rad/step).
The ACT base always owns the grippers.

Real-rig outcome referenced by the project repo: full 750-step run reaches status=0 with SmolVLA-dominated segments at authority ~0.7–0.9.

Files

Same layout as a LeRobot exported pretrained_model/: config.json, model.safetensors, policy pre/post-processors with their normalizer/unnormalizer, and the tokenizer/ from the backbone. Load it with the SmolVLA fork of LeRobot in the GitHub repository above.

Task: Reinforcement Learning
Library: lerobot
License: apache-2.0
Created: Sep 2, 2026
Updated: Sep 2, 2026