Karma: realtime control and policy runtime for YAM.

Karma runs config-driven, multi-process realtime sessions for YAM follower arms, YAM/GELLO-style leader arms, cameras, simulation, and policy agents.

git clone --recurse-submodules https://github.com/SRA-VJTI/karma.git
cd karma
uv sync

Optional camera/sim extras:

uv sync --extra sensors
uv sync --extra realsense
uv sync --extra mjlab-sim

uv run krm session configs/bimanual_yam_leader.yaml
uv run krm session configs/sim_gello_teleop.yaml
uv run krm session configs/bimanual_openpi_policy_xdof_hq.yaml --save-root recordings

Disable the TUI if needed:

uv run krm session configs/sim_dummy.yaml --no-tui

Replay a recorded sim episode:

uv run krm replay recordings/<date>/<episode_dir>

• configs/bimanual_yam_leader.yaml — YAM active leader arms → YAM followers.
• configs/bimanual_gello_teleop.yaml — serial GELLO leaders → YAM followers.
• configs/bimanual_gello_teleop_bair_autolab.yaml — lab camera variant.
• configs/bimanual_passive_gello_teleop_xdof_hq.yaml — SocketCAN passive leaders.
• configs/bimanual_act_policy_xdof_hq.yaml — remote ACT policy deployment.
• configs/bimanual_openpi_policy_xdof_hq.yaml — OpenPI policy deployment.
• configs/bimanual_molmoact2.yaml — MolmoAct2 policy deployment.
• configs/sim_dummy.yaml — autonomous dummy leaders into the YAM sim.
• configs/sim_gello_teleop.yaml — leader teleop into the YAM sim.

Robot hardware configs live directly under robot_configs/.

Use simple bus node names:

• followers: left, right
• leaders: leader_left, leader_right
• simulation: sim

Examples:

- type: RobotNode
  name: left
  cmd_topic: leader_left/joint_pos

state_topics:
  left: left/joint_state
  right: right/joint_state

- type: XdofSimNode
  name: sim
  cmd_topics:
    left: leader_left/joint_pos
    right: leader_right/joint_pos