ACT β€” flexnsense skill 1 (right index β†’ right middle finger)

ACT policy (LeRobot 0.7.1) for a NormaCore ElRobot follower arm (7 joints + gripper, 8 Feetech ST3215 servos, IDs 1..8).

  • Task: move device from right index to right middle finger
  • Checkpoint: step 030000 of 30000 (batch size 32)
  • Training data: JustoDario/flexnsense_skill1_lerobot β€” 43 teleoperated episodes, 8465 frames, 10 Hz
  • chunk_size / n_action_steps: 30 / 15 (3 s / 1.5 s at 10 Hz)

Inputs / outputs

Feature Shape Meaning
observation.state (8,) NormaCore position_norm in [0, 1], ordered by motor ID 1..8 (ID 8 = gripper)
observation.images.side (3, 224, 224) RGB camera
observation.images.gripper (3, 224, 224) RGB camera
observation.images.table (3, 224, 224) RGB camera
action (8,) NormaCore goal_norm in [0, 1], same order

To drive the servos, map each action to raw ticks with the per-joint calibration range: ticks = range_min + goal_norm * (range_max - range_min).

Camera views: side (side view of the arm), gripper (gripper camera), table (table-level view of the hand). NormaCore Station does not label the images of the inference/normvla queue and their order can change between restarts, so identify each camera by content before feeding the policy.

Usage

from lerobot.policies.act.modeling_act import ACTPolicy
from lerobot.policies.factory import make_pre_post_processors

repo = "JustoDario/flexnsense_act_skill1_rightindex2middle"
policy = ACTPolicy.from_pretrained(repo)
pre, post = make_pre_post_processors(policy.config, pretrained_path=repo)

The normalization statistics are stored in the processor files of this repo. Not evaluated on the real robot at upload time.

Downloads last month
34
Safetensors
Model size
51.6M params
Tensor type
F32
Β·
Video Preview
loading