Miical/pi05-libero-spatial-sft-step-100

PI0.5 LIBERO Spatial SFT — Step 100

This is a deliberately undertrained PI0.5 checkpoint fine-tuned on LIBERO Spatial. It is intended as a reproducible starting policy for measuring how much downstream reinforcement learning improves task success, rather than as a fully trained SFT release.

The checkpoint is an upstream-native, merged PI0.5 policy exported by verl-vla. It does not require a verl-vla checkpoint wrapper for inference.

Lineage

• Base model: Miical/pi05-base
• Training dataset: lerobot/libero_spatial_image
• Benchmark: LIBERO Spatial, all 10 tasks
• Checkpoint: optimizer step 100

Training configuration

Setting	Value
Training method	Supervised fine-tuning (SFT)
Optimizer steps	100
Dataset size	52,970 frame-level samples, 432 episodes
Samples consumed	25,600
Effective epochs	approximately 0.49
Global / mini-batch size	256
Micro-batch size	16
GPUs	8 x NVIDIA H20
Distributed strategy	FSDP2
Model dtype	bfloat16
Learning rate	1e-4
Weight decay	1e-5
LR warmup ratio	0.05 (10 steps in the 200-step run)
Action horizon	10
DataLoader workers	8
Seed	42

One complete pass over this dataset is 206 optimizer steps with the configured batch size. This checkpoint was saved at step 100, before completing the first epoch. It therefore provides a capable but intentionally non-saturated policy for subsequent RL experiments.

LIBERO Spatial evaluation

Evaluation used all 10 LIBERO Spatial tasks with 10 trials per task, a maximum of 256 environment steps per trial, and the same verl-vla evaluation workflow used for later SFT checkpoints.

Task	Instruction	Successes	Success rate
0	Pick up the black bowl between the plate and the ramekin and place it on the plate	7 / 10	70%
1	Pick up the black bowl next to the ramekin and place it on the plate	8 / 10	80%
2	Pick up the black bowl from table center and place it on the plate	6 / 10	60%
3	Pick up the black bowl on the cookie box and place it on the plate	8 / 10	80%
4	Pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate	3 / 10	30%
5	Pick up the black bowl on the ramekin and place it on the plate	6 / 10	60%
6	Pick up the black bowl next to the cookie box and place it on the plate	6 / 10	60%
7	Pick up the black bowl on the stove and place it on the plate	8 / 10	80%
8	Pick up the black bowl next to the plate and place it on the plate	5 / 10	50%
9	Pick up the black bowl on the wooden cabinet and place it on the plate	3 / 10	30%
Overall	All tasks	60 / 100	60%

These results are a single 100-trajectory benchmark run. Per-task estimates use only 10 trials and should not be interpreted as precise population rates.

Intended use

This model is intended for experiments that start from an incompletely trained SFT policy and evaluate whether reinforcement learning can improve success rate, especially on the weaker LIBERO Spatial tasks. For a high-performing SFT baseline or deployment candidate, use a later and more thoroughly evaluated checkpoint instead.

Loading

from giga_models import PI0Policy

policy = PI0Policy.from_pretrained(
    "Miical/pi05-libero-spatial-sft-step-100"
)

With verl-vla, set the model path directly:

cluster:
  actor_rollout_ref:
    model:
      path: Miical/pi05-libero-spatial-sft-step-100
      adapter:
        embodiment: libero

License and attribution

This fine-tuned checkpoint inherits the usage and redistribution constraints of the original PI0.5 checkpoint and its underlying components. The PyTorch conversion and verl-vla integration code have their own respective licenses; consult the base model card and upstream projects before redistribution or commercial use.