PiStar0.6 Value Function — Insert Carrot v2+v3
PiStar0.6 value model trained from scratch on the round-2 merged dataset
AlvinAi/insert_carrot_into_the_hole_hil_v2_v3_merged: the cleaned
insert_carrot_into_the_hole_hil_v2_r1 plus insert_carrot_into_the_hole_hil_v3_r1
with long initial static segments trimmed (motion threshold 0.01, 5 consecutive
moving frames, 1 s of static context kept).
Training
- Frames: 126,443
- Episodes: 254 (83 from v2 r1, 171 from v3 r1)
- Cameras:
observation.images.front,observation.images.side - GPUs: 1
- Batch size: 64
- Optimizer updates: 10,000 (same sample budget as r1: 2,500 updates x 256)
- Precision: bfloat16
- Gradient checkpointing: enabled
- Warmup steps: 1,000
- Scheduler decay steps: 10,000
- Checkpoints: steps 2,500, 5,000, 7,500 and 10,000
The final checkpoint contains the value model, configuration, processor definitions, and normalization statistics. Training optimizer and RNG state remain local and are not uploaded.
- Downloads last month
- 4