pi05_telearms_v6_lora

π₀.₅ (lerobot/pi05_base) LoRA fine-tune for the task "Red Left Of Blue Placement" on a simulated bimanual YAM, merged into a full model.safetensors, so it loads like a normal checkpoint.

Evaluation

Closed-loop simulation, one task: 49/50 episodes reached success (98%, Wilson 95% CI 89.5–99.6%). 47/50 (94%, CI 83.8–97.9%) still met the success condition at the end of the episode.

Read this as performance on the training image distribution, not robustness or sim-to-real: observations use the same rig cameras and rig DR v7 rendering as the training data. That required a local patch to training/evaluate_pi.py (TELEARMS_EVAL_RIG=1) that is not on the dr-rig branch, so these numbers are not comparable with the stock eval's cameras. No real-robot test has been run.

• Seeds 900000001–900000050, one policy sample seed per episode, 1,800-step cap (60 s at 30 Hz), replanning every 25 steps. Start states come from the task's reset distribution, not the validation episodes.
• "Success" means the environment's success condition was met at any step. The policy stops at that point and its last command is held for 150 steps; the arms do not retreat.
• First success: median step 396 (13 s), slowest step 1,647.
• The one failure, seed 900000010: red moved toward its goal but never got far enough left of blue (best y separation -0.021 m, needed +0.03 m), and the cubes stopped moving after step 1,408.
• Seeds 900000014 and 900000028 met the condition for only 2 frames and were false at the end, with the cubes on the table and correctly separated. The cause was not diagnosed.
• The first run crashed on a 300 s simulator timeout during seed 900000036, before scoring it. That seed and the remaining ones were run again with a 900 s timeout, and it did not recur.

Data

albagon/telearms-yam-red-left-blue-v6-rig-dr-v7-tar (revision b86b47af): 9,842 episodes from 4,921 demos, 4,075,628 frames at 30 Hz, images with rig DR profile 7.

8 episodes were dropped by the length/motion filter and 994 held out for validation, leaving 3,665,336 training frames. The run saw 50.0% of those.

Known data risk (from the dataset's README): rendered on a GPU whose VRAM showed bit errors. Validation checks files and data rows, not image pixels, so isolated frames may be silently corrupted.

Training

• Code: helen9975/RoboEval-telearms branch dr-rig @ ea3e0a2 (tools/training_cluster/run_lora_helen_4h100.sh) plus a local change to training/train_pi.py (NaN guards), LeRobot 0.6.1 @ e40b58a8 with the repo's patch.
• LoRA r=128, alpha=128 on the PaliGemma language model and Gemma expert layers; action_in_proj, action_out_proj, time_mlp_in and time_mlp_out trained in full; SigLIP frozen.
• 28,636 steps at batch 16/GPU on 4 × A100 80GB (global 64), bf16.
• Peak LR 2e-4 for adapters and 5e-5 for full modules. These are the earlier run's rates at global batch 16 scaled by sqrt(64/16); the recipe's authors mark that scaling as unverified for LoRA.
• State dropout 0.3, state noise 0.05, image augmentation on, per-task prompt variants.
• Training loss near the end ≈ 0.032; held-out loss over the last 3,500 steps 0.032–0.052 (noisy: each eval uses few samples).
• The job was preempted once and resumed from step 24,344. All weights checked finite after merging.