pi0.5 Delta-flow — R1Lite microwave, final epoch 60
This is the completed 60-epoch, 18,060-update model, fully fine-tuned from the original pi0.5 base. It is not LoRA and is not resumed from the other method. All 3,616,757,520 parameters were trainable. The complete FP32 model.safetensors is approximately 14.5 GB; no separate base-model download is required for inference.
The older epoch-42 release is preserved at the epoch-042 tag and immutable revision 7deba64677719d0c70df6079d15e7e8a7e9c0ecd. The default branch now contains the final model.
Instruction: Place the bowl inside the microwave and close the door. This policy predicts actions, not future images.
Training report
Both runs used the same 34 demonstrations / 14,432 frames at 15 FPS, pinned dataset revision 26398706f75f0695572fc789ec19b0aa180bf3e7, the same task-specific normalization, original base checkpoint, and random seed 42. Each used 8 A100 80-GB GPUs, batch 6 per GPU / global batch 48, 301 updates per epoch, FP32 parameters and AdamW states with BF16 autocast, gradient checkpointing, 903 warmup updates and cosine learning-rate decay from 5e-5 to 1e-6. AdamW betas are (0.9, 0.95), weight decay 0.01 and gradient clipping 1.0.
| Final epoch-60 logged mean | Delta-flow | Gaussian baseline |
|---|---|---|
| FM loss | 0.000571062 | 0.001642929 |
| IC loss (unweighted) | 0.003339722 | Not used |
| Total loss | 0.002240923 | 0.001642929 |
| Formal training wall time | 12 h 56 min | 7 h 36 min |
Each epoch-60 mean averages 31 logged current-batch means, sampled every 10 updates; it is not an exact average over all training updates. The total objectives differ and the FM velocity targets use different sources. Neither total-loss nor FM-loss ratios are policy-success ratios. All demonstrations were used for training; no held-out or quantified physical-robot success rate is asserted here.
See training report, summary JSON, per-update logs as CSV, epoch summaries, and PDF. Raw fluctuations and causal smoothing are shown separately. These are single-seed optimization diagnostics, not replicated-run confidence intervals.
Source and objectives
Action source: normalized past 16 actions plus Gaussian noise with standard deviation 0.5 on arm joints only. Grippers 6/13 and padding 14:32 receive no source noise. Training uses FM + 0.5 × one-step L1 endpoint consistency.
Offline inference
Use Python 3.11 in a new, dedicated virtual environment, and sufficient RAM/VRAM for the full model plus activations. CPU uses FP32; BF16-capable CUDA uses FP32 parameters with BF16 autocast. Setup only installs the bundled compatibility overlay into the active dedicated virtual environment. No model/tokenizer downloads occur during inference.
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip huggingface_hub
hf download ZUO66/pi05-r1lite-microwave --local-dir pi05-microwave-delta
cd pi05-microwave-delta
python -m pip install -r requirements_inference.txt
python setup_runtime.py
python infer_microwave.py --release . \
--head example/head.png --left-wrist example/left_wrist.png \
--right-wrist example/right_wrist.png --inputs example/inputs.npz \
--output prediction.npz --device cuda --seed 42 --num-steps 1
Replace --device cuda with --device cpu for CPU use. Delta-flow requires one Euler update. The Gaussian baseline defaults to ten updates and also supports --num-steps 1; use the same observation and seed for an offline solver-budget diagnostic. Omit the seed for fresh source noise on each call.
Inputs are synchronized RGB head/left-wrist/right-wrist images as three separate views, each resized with official aspect-preserving zero padding to 224 × 224, and raw current 14D state. State/action order is [left_arm6, left_gripper, right_arm6, right_gripper]. Actions are absolute joint targets, not Cartesian deltas. The internal representation has 32 dimensions, but the helper returns 16 × 14 raw-scale predictions using this release's norm_stats.json. Do not use the flower task's 7D interface or another task's statistics.
Delta inputs.npz requires proprio[14], past_actions[16,14] in chronological order strictly before the current observation, and optionally past_action_valid_mask[16]. A short history is left-padded by repeating its earliest action, with false validity for padded rows. With no genuine history, a safe initial command repeated 16 times and --allow-empty-history is an explicit, unvalidated deployment initialization. Future states/actions must never be used as inputs.
The supplied input is a training-set example (episode 0, frame 134). Validation checks strict final-checkpoint loading, exact cached state tokens and finite offline predictions, not task success. The helper never sends robot commands. No CUDA latency benchmark is claimed from CPU validation.
Contents, limitations and safety
The release includes final weights, sanitized training configuration, normalization, local tokenizer, exact inference runtime, an offline example, training curves/source CSVs, and checksummed manifests. Optimizer/RNG state, full demonstrations, caches and credentials are not uploaded. The original server training checkpoints are unchanged.
Before real execution, independently confirm camera orientation, RGB preprocessing, joint/gripper order and units, current-state synchronization, absolute-control interface, robot limits and collision safety, action clipping, history buffering where applicable, and how many actions are executed before replanning. A predicted 16-step chunk is not an instruction to execute all 16 blindly. Dataset timing is 15 FPS; deployment timing requires separate verification.
The bundled OpenPI code license is retained as LICENSE.openpi. No new license over upstream model weights, tokenizer or dataset is asserted; applicable upstream terms remain in effect.