Aikwed/pistar06_insert_carrot_hil_r1_value

Pi*0.6 Value Function — Insert Carrot

This repository contains the final step-2000 Pi*0.6 value-function checkpoint trained for the task "insert carrot into the hole". It was trained with the MINT-SJTU/Evo-RL pipeline on human-in-the-loop robot demonstrations.

This is a value model, not an action policy. Its intended role is to estimate per-frame value/return-to-go signals that Evo-RL can convert into n-step advantages and binary Advantage-Conditioned Policy (ACP) indicators.

Training data

• Dataset: AlvinAi/insert_carrot_into_the_hole_hil
• Frames: 40,353
• Episodes: 83
• Task: insert carrot into the hole
• Observation cameras: observation.images.front and observation.images.side
• Robot state: observation.state, 6 dimensions

Model configuration

• Value model type: pistar06
• Vision backbone: google/siglip-so400m-patch14-384
• Language backbone: google/gemma-3-270m
• Value support: 201 bins over [-1.0, 0.0]
• Training dtype: bfloat16
• Vision and language backbones were fine-tuned
• Gradient checkpointing: enabled

Training configuration

Parameter	Value
GPUs	4
Batch size per GPU	64
Global batch size	256
Optimizer steps	2,000
Approximate dataset passes	12.7
Optimizer	AdamW
Peak learning rate	5e-5
Warmup steps	125
Scheduler decay steps	2,000
Final checkpoint	step 2,000

The final W&B training summary reported train/loss=0.5956 and train/value_mae=0.00191. These are training metrics, not held-out evaluation or robot rollout success rates.

Value inference and ACP annotation

Run from an Evo-RL environment:

lerobot-value-infer \
  --dataset.repo_id=AlvinAi/insert_carrot_into_the_hole_hil \
  --inference.checkpoint_path=Aikwed/pistart0.6-insert-carrot \
  --runtime.device=cuda \
  --runtime.batch_size=64 \
  --acp.enable=true \
  --acp.n_step=50 \
  --acp.positive_ratio=0.3 \
  --acp.value_field=complementary_info.value_pistar06 \
  --acp.advantage_field=complementary_info.advantage_pistar06 \
  --acp.indicator_field=complementary_info.acp_indicator_pistar06 \
  --output_dir=outputs/value_infer/pistar06_insert_carrot

The checkpoint includes the model weights, model/training configuration, processor definitions, and normalization statistics. Optimizer, scheduler, and RNG training states are not included in this Hub repository.

Limitations

• The model is task- and embodiment-specific and has not been validated outside the insert-carrot dataset distribution.
• Camera keys, image shapes, robot-state layout, and preprocessing must match the saved configuration.
• Low offline training loss does not by itself establish calibrated values or improved closed-loop robot success; downstream rollout evaluation is required.