Skip to content

Latest commit

 

History

History
55 lines (34 loc) · 1.88 KB

File metadata and controls

55 lines (34 loc) · 1.88 KB

MaskACT

  • pyqueat 형태로 변경하기 : scripts > rosbag to lerobot

Train

  • model 변경하기 web 으로 못돌리면 그냥 offline으로 가자 (2h)
  • 학습시간 4090 train 하루
  • 시간되면 —> 일반 ACT

inference time

  • inference time에 mask 이미지 만들어서 보내는 코드 (확인) (60분)
  • /tf 에서 /endpose 보낼 코드 필요 (60분이면 함) : ==> inference_time_topic 으로 통일
  • inference code 짜거나 변경 —> 이것도 gr00t 처럼 offline으로 가기 (2h)

inference 해보기

  • 지금 코드로 inference in the sim
  • real test

RL

  • 강화학습 테스트 : 허나만 나오게 해서 테스트

VLA

  • Gr00t팔 뒤로가는 이유 -> 하나만 일단 보내보기 /leader 통해서 offset이 있는지 보기 (@2026-01-07)

  • 4090으로 학습 인퍼런스 빨리 갈 수 있게 만드는 방법

    • multi step mode test : 잘 안됨

    기존 방법 (기본값)

    ros2 run ... --ros-args -p action_processor_mode:=original

    새로운 방법 (속도 향상)

    ros2 run ... --ros-args
    -p action_processor_mode:=multi_step
    -p action_scale:=2.0
    -p action_lookahead_steps:=3
    -p max_joint_delta:=0.5

데이터

  • 추가 모음 [0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 27 28 29 30 31 32 34 35 36 37 38 39 40 41]

==> 애초에 maskact 추가하는 코드에서 state를 그냥 다 만들었어야 했네 image 별로 만드는게 아니라

inference 돌려보기 위함

ros2 bag play /workspace/dkim_result_3/100/converted_20260107_143050_fdf2/ -l --clock --exclude-topics /leader/joint_trajectory_command_broadcaster_right/joint_trajectory /leader/joint_trajectory_command_broadcaster_left/joint_trajectory /mf_perception/ch0_pruning_indiocator

  • gr00t 이미지 구워진 상태에서 컷밋