README.md

ARX real-data adapter

arx_lerobot_adapter.py converts ARX LeRobot joint data into TCP and flange poses expressed in each arm's zero-flange coordinate frame:

• left_flange_zero
• right_flange_zero

It reads joint values from observation.state by default, runs FK against the ARX MuJoCo scene, and does not modify the source dataset.

Each arm/frame stores:

• the six raw joint positions;
• T_flange_in_output_frame;
• T_tcp_in_output_frame;
• the original gripper scalar as gripper.raw;
• gripper.continuous in [0, 1], with fully open 0 and fully closed 1;
• gripper.binary, with open 0 and closed 1;
• the corresponding raw/continuous/binary action values when available.

The default ARX raw calibration is open -3.4, closed 0.1, with a binary threshold of -2.6. Values greater than or equal to -2.6 are classified as closed. The threshold is configurable from the CLI and is recorded in every output file. With the current recordings, the expected state sequence begins with closed, followed by open, closed, and open.

Convert one episode

PY=/home/xule/miniconda3/envs/lifego/bin/python

$PY -m real_data.arx_lerobot_adapter \
  --input /home/xule/le_ws/Data_TATE/stack_cube_arx/data/chunk-000/file-000.parquet \
  --out outputs/arx_real_flange/stack_cube_arx_ep000

The episode is written as episode_000000.json, accompanied by manifest.json.

Convert the complete dataset

PY=/home/xule/miniconda3/envs/lifego/bin/python

$PY -m real_data.arx_lerobot_adapter \
  --input /home/xule/le_ws/Data_TATE/stack_cube_arx \
  --out outputs/arx_real_flange/stack_cube_arx

To test another threshold without overwriting the first result, use a new output directory:

$PY -m real_data.arx_lerobot_adapter \
  --input /home/xule/le_ws/Data_TATE/stack_cube_arx \
  --out outputs/arx_real_flange/stack_cube_arx_threshold_m2p7 \
  --gripper-binary-threshold-raw -2.7