pi05_telearms_frontback_lora

π₀.₅ (lerobot/pi05_base) LoRA fine-tune for the task "Front To Back Cube Line" on a simulated bimanual YAM. Checkpoints are merged into a full model.safetensors, so each loads like a normal checkpoint, and live in merged/<step>/.

Status: training is still running. Uploaded so far:

Folder	Step	Share of the run	Held-out loss near this step
merged/020766/	20,766	30%	0.047–0.053 (steps 20,000–21,000)
merged/034610/	34,610	50% (halfway)	0.036–0.043 (steps 33,500–34,500)
Held-out loss is noisy because each eval uses few samples.

Evaluation: not yet evaluated, in simulation or on a robot.

A checkpoint from partway through has the learning rate of that point in the schedule, which has not decayed yet, so it is not equivalent to a finished shorter run.

Data

albagon/telearms-yam-front-to-back-v6-rig-dr-v7-tar (revision 071293be): 40,146 episodes from 20,073 demos, 39,254,588 frames at 30 Hz, images with rig DR profile 7.

20 episodes were dropped by the length/motion filter and 3,964 held out for validation, leaving 35,435,642 training frames. The full run is set to see 25% of those.

Training

• Code: helen9975/RoboEval-telearms branch dr-rig @ 455bd9f (tools/training_cluster/run_lora_frontback.sh, config 25pct_8gpu) plus a local change to training/train_pi.py (NaN guards).
• LoRA r=128, alpha=128; SigLIP frozen.
• 69,212 steps planned at batch 16/GPU on 8 × A100 80GB (global 128), saving every 3,461 steps.
• Peak LR 2.83e-4 for adapters and 7.07e-5 for full modules: the earlier run's rates at global batch 16 scaled by sqrt(128/16). The recipe's authors mark that scaling as unverified for LoRA.