flux3-chess-lora-kitchenconv

LoRA adapters for FLUX 3 Action SO-101 (black-forest-labs/flux-3-action-so101 @ c9e13b2aca6a0a472b3ea03fd90cff32d3e85849), trained with LeRobot's examples/flux3/lora.json recipe on one SO-101 chess move (horse c3 → d5).

• Dataset: niru-5/chess_dataset_c3_6cm_to_d5_10cm_20261017_145741 @ 9b6b38719ab08509bdf4ce169a4de028573199de; last 0.2 of episodes held out (episodes 24–30); 125 steps at batch 2 per GPU × accum 1 (4 × H100, DDP), seed 42.
• Repo root = step 125 (lowest in-training eval loss so far), EMA weights. Every finished checkpoint is under checkpoints/<step>/ with pretrained_model/ (raw) and pretrained_model_ema/ (EMA). Optimizer state is not uploaded.
• Encoders: black-forest-labs/flux-3-action-base @ 62878e2925e59b7a89ec14463ce89932624c490d (video VAE, text encoder).
• Use is subject to the license of the base FLUX 3 Action models.
• Joint convention: "kitchen" (the convention of VoicAndrei/so100_kitchen, whose statistics the released checkpoint ships). Trained on a local copy of the dataset with shoulder_lift' = 86 − shoulder_lift and elbow_flex' = elbow_flex + 77 applied to both observation.state and action; other joints unchanged; the checkpoint's own normalization statistics are kept. Rollout must apply the same map to the robot state and the command history before the model, and the inverse (lift = 86 − lift', elbow = elbow' − 77) to the model's commands before joint limits and the arm. Do not combine with a --normalization override. The offsets are estimates from rest poses (±5–10°). Offline MAE is in degrees and comparable with the other repos (the map does not change error magnitudes); the "base" row here is the base model on the converted data.

Checkpoints

Eval loss: LeRobot's in-training flow-matching loss on the held-out episodes, computed on the raw weights with random noise (differences of a few percent are noise). Offline MAE: teacher-forced action error of the EMA weights on the same held-out episodes (88 windows, stride 32, one sample each): predicted vs recorded commands over the executed horizon (first 32 of 42 steps); arm = mean of the 5 arm joints in degrees, gripper in %. Lower is better.

step	eval loss (raw)	offline MAE arm ° (EMA)	gripper %	folder
125	0.1187	7.91	3.88	checkpoints/000125/ root
base so101	–	9.98	4.09	black-forest-labs/flux-3-action-so101

Per joint, root checkpoint (step 125) vs base

joint	base so101	step 125
shoulder_pan	9.43	7.64
shoulder_lift	8.45	6.65
elbow_flex	11.47	9.59
wrist_flex	13.74	10.27
wrist_roll	6.83	5.40
gripper	4.09	3.88

Load

from peft import PeftConfig, PeftModel
from lerobot.policies.flux3.modeling_flux3 import Flux3Policy

repo, sub = "Gawtam/flux3-chess-lora-kitchenconv", None  # or sub="checkpoints/000250/pretrained_model_ema"
cfg = PeftConfig.from_pretrained(repo, subfolder=sub)
policy = PeftModel.from_pretrained(Flux3Policy.from_pretrained(cfg.base_model_name_or_path), repo, subfolder=sub)

Run: lora_c3d5_kitchenconv.