pi05_franka_haply_abscartesianpos

π₀.₅ fine-tuned from lerobot/pi05_base on Ameyapores/franka_haply_joint_delta, predicting absolute end-effector position and gripper.

Checkpoint taken at the step with the lowest eval loss: step 1,750, eval_loss 0.1120.

Action space

The policy outputs a 4-dimensional action, and the order is part of the contract:

index	meaning	units
0	position_x	metres, absolute workspace coordinate
1	position_y	metres, absolute
2	position_z	metres, absolute
3	gripper	normalised open/close

These are absolute positions, not displacements. Feeding the output to a controller that expects a delta will drive the arm to the workspace origin. Observed ranges in the training data are x ∈ [0.35, 0.66], y ∈ [−0.16, 0.19], z ∈ [0.13, 0.49] m — a Franka envelope, not increments.

The source column is action_cartesian_absolute. Note the dataset's default action column holds joint deltas and action_cartesian holds delta xyz despite the similar name; this model was trained on neither.

Observations

Three 224×224 RGB cameras, each a single frame — π₀.₅ uses no observation history (n_obs_steps=1), so all temporal structure comes from predicting a chunk forward:

• observation.images.base_0_rgb
• observation.images.base_1_rgb
• observation.images.left_wrist_0_rgb

plus an 8-dim observation.state and a language prompt. The dataset is recorded at 20 fps, so chunk_size=50 is 2.5 s of motion.

Training

base	lerobot/pi05_base
dataset	Ameyapores/franka_haply_joint_delta, 94 episodes / 50,861 frames @ 20 fps
split	84 train / 10 eval (last 10 episodes, eval_split=0.1)
action scheme	action_cartesian_absolute → 4-dim action
chunk_size / n_action_steps	50 / 50
batch / LR	256 global, 2.5e-5, cosine
precision	bfloat16, gradient checkpointing
normalisation	quantile (state and action)
trainable	--train_expert_only — SigLIP and Gemma-2B frozen; 693M of 4.14B trainable
schedule	3,500 steps (19.2 epochs) on 8×MI300X, eval every 250

eval_loss is the flow-matching objective on held-out episodes — not a success rate. The holdout is 10 episodes, so differences of a few percent are not resolved.

Comparison

All arms below share this base, chunk, optimiser and the same 84/10 holdout, so the losses are directly comparable. Only the action encoding differs.

action encoding	dims	best eval
absolute xyz + gripper (this model)	4	0.1120
absolute xyz + gripper, sliced from the older 8-dim column	4	0.1124
xyz + quaternion + gripper	8	0.1438
absolute joint positions	8	0.1454
joint deltas	8	0.1977
delta xyz + gripper	4	0.2122

The last row is the informative one: identical in every respect except absolute versus incremental position, and 47% worse. With a single observation frame, a delta is a rate regressed from a snapshot, while an absolute target is directly readable from the current image and state. Absolute encodings occupy the top of this table; incremental ones the bottom.

Usage

from lerobot.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("arkojit1/pi05_franka_haply_abscartesianpos")

Caveats

• One seed, one task, one operator. The 0.0004 gap to the sliced-column variant is noise, not a ranking.
• Trained on Haply-teleoperated demonstrations of a single language task; no claim is made about transfer to other tasks or robots.
• Evaluated by held-out loss only. No rollout success rate has been measured.