pi0.5 folding — training-time RTC, max delay 10 (step 3000)

π₀.₅ fine-tuned with training-time real-time chunking so that inference can run with --inference.rtc.mode=trained: during training the last d actions of the previous chunk are held clean and the remainder is denoised, with d drawn uniformly from 0…10 per example. The model therefore learns to complete a chunk whose prefix is already committed, which is what makes delay-compensated inference sound rather than approximate.

This is step 3000 of a planned 5000-step run. The interactive session hosting the job was lost at around step 3800, so this is the last checkpoint written and the cosine schedule had not annealed (LR was 1.2e-05 at the point of loss). Loss was flat from step 2100 onward, so the missing 2000 steps are unlikely to matter much for behaviour, but it is not a fully decayed model.

base checkpoint	lerobot-data-collection/folding_final @ 695abe40dbf3aac04efda59c1501d748681fa0fb
dataset	lerobot-data-collection/level2_final_quality3_t_0_hil_data_c @ 2496db53d330c360f910d095e13698d968c56fc6 — 1319 episodes, 3,414,338 frames
code	huggingface/lerobot @ 4dbb965455ed8c260e5ba947e55558b4d4c14b5d (training-time RTC, PR #4056)
rtc_training_max_delay	10
chunk_size	30
relative actions	enabled (relative_exclude_joints: ['gripper'])
hardware	8×H100 80GB, one node, 45.96 GB per GPU
batch	12 per GPU × 8 = 96
optimizer	AdamW, LR 3.75e-05, weight decay 0.01, cosine decay with 500 warm-up steps over 5000
throughput	1.22 s/step, ~78 samples/s

Training command

Run under accelerate on a single 8×GPU node. --policy.path points at a local copy of the base snapshot, because folding_final predates the delta_actions_processor → relative_actions_processor registry rename and that one string has to be patched in policy_preprocessor.json before it will load.

accelerate launch \
  --num_machines=1 \
  --num_processes=8 \
  --main_process_port=29827 \
  lerobot-train \
  --policy.path=/path/to/folding_final_snapshot_copy \
  --policy.rtc_training_max_delay=10 \
  --policy.device=cuda \
  --policy.dtype=bfloat16 \
  --policy.gradient_checkpointing=true \
  --policy.compile_model=false \
  --policy.push_to_hub=false \
  --policy.scheduler_warmup_steps=500 \
  --policy.scheduler_decay_steps=5000 \
  --dataset.repo_id=lerobot-data-collection/level2_final_quality3_t_0_hil_data_c \
  --dataset.root=/path/to/level2_final_quality3_t_0_hil_data_c_snapshot \
  --dataset.use_imagenet_stats=false \
  --output_dir=/path/to/output \
  --job_name=folding-final-rtc-d10-5k \
  --batch_size=12 \
  --steps=5000 \
  --num_workers=8 \
  --save_freq=1000 \
  --log_freq=100 \
  --seed=1000 \
  --wandb.enable=true \
  --wandb.project=folding-rtc \
  --wandb.disable_artifact=true

--policy.compile_model=false is required rather than cosmetic: the base checkpoint carries compile_model: true with compile_mode: max-autotune, and triton fails to lower it (KeyError: 'cubin') on this cluster. --policy.dtype=bfloat16 with no AMP keeps AdamW state in bf16, which is why a full 4.1B-parameter fine-tune fits in 46 GB.

Loss

step	100	600	1100	1600	2100	2600	3100	3600
loss	0.428	0.036	0.029	0.026	0.026	0.023	0.023	0.021
grad norm	5.71	0.38	0.27	0.25	0.24	0.25	0.25	0.25

Inference

lerobot-rollout \
  --strategy.type=base \
  --policy.path=nepyope/rtc \
  --inference.type=rtc \
  --inference.rtc.mode=trained \
  --inference.rtc.execution_horizon=10 \
  --inference.queue_threshold=10 \
  --robot.type=bi_openarm_follower \
  --task="Fold the T-shirt properly" \
  --fps=30 \
  --interpolation_multiplier=3

execution_horizon must satisfy rtc_training_max_delay <= horizon <= chunk_size - rtc_training_max_delay, so the valid window here is 10 to 20, and queue_threshold has the same lower bound of 10. Measured rollout delay on the reference hardware is ~7 frames, so 10 is the reactive end of the range; raise it if calls overrun and chunks get discarded. A one-off spike past the trained delay discards that chunk and retries — only a sustained overrun is fatal.

The relative-actions processor must stay enabled: RTC inference requires an enabled RelativeActionsProcessorStep to reconstruct absolute targets from the committed prefix.