Fine-Tune Pi0.5 with Language Table Data

Task Performance of R3 + Diff Checkpoints

OpenPI Evaluation Configuration

• Checkpoint: pi05_language_table_full/language_table_full_pytorch_bs128/35000
• Action horizon: 10
• Actions executed per OpenPI query: 5
• VLM chunk length: 20
• Maximum step count: 20

Task index	Task name	Fine-tuned Pi0.5	Pre-trained ckpt
0	group_color	85.2%	65.8%
1	line_color	46.4%	32.4%
2	T	13.7%	9.9%
3	V	79.9%	69.2%
4	rect	10.0%	6.0%
5	gris	60.6%	47.8%
6	mid	64.6%	51.0%
7	L	53.5%	35.2%
8	clear_quarter	92.5%	93.8%
9	iip	43.0%	35.5%
10	iV	68.6%	57.5%
11	iL	49.9%	37.2%
12	dline	43.7%	30.9%
14	clear_half	78.0%	74.5%

Single-Instruction Evaluation of Fine-Tuned Pi0.5

Evaluation Results

Eight models were evaluated across five single-instruction Language Table tasks, with 50 episodes per task (250 episodes per model).

Model	Abs. loc.	Block-to-block	Rel. loc.	Block rel. loc.	Separate	Overall
bc_resnet_sim_checkpoint_955000	78%	90%	58%	56%	92%	74.8%
language_table_full_pytorch_bs128_25000_exe1	80%	72%	74%	80%	100%	81.2%
language_table_full_pytorch_bs128_30000_exe1	82%	90%	82%	74%	100%	85.6%
language_table_full_pytorch_bs128_35000_exe1	84%	90%	78%	88%	100%	88.0%
language_table_full_pytorch_bs128_35000_exe10	80%	90%	74%	68%	98%	82.0%
language_table_full_pytorch_bs128_35000_exe5	78%	84%	86%	90%	98%	87.2%
language_table_full_pytorch_bs128_40000_exe1	78%	80%	66%	72%	100%	79.2%
language_table_full_pytorch_bs128_50000_exe1	82%	88%	66%	76%	100%	82.4%

Best overall: language_table_full_pytorch_bs128_35000_exe1 with 220/250 successes (88.0%).

Analysis

For execution-length-1 checkpoint variants, the 35k checkpoint performs best.

At the same 35k checkpoint, execution lengths 1 and 5 are comparable. Execution length 10 drops significantly, primarily on blocktoblockrelativelocation.