Template updates

MolmoAct2 SO-100/101

molmoact2-so101 · policy template by Ai2 · SO-100/101 · 30 Hz · L4. Each version below is fixed once published, newest first.

seq init molmoact2-so101 gives the stable version, seq init molmoact2-so101@beta the beta, and seq init molmoact2-so101@v<N> any version below that is not withdrawn. A deployment made from this template names the version it runs in seq policy versions and on its page in the console; seq policy upgrade moves it to the newer stable one.

Performance

No runs yet.

Versions

  1. v6stablebeta

    As v5; policy.py says, as the README does, how a worker's CPU and memory are billed.

    Needs sequence-ai 0.56.0 or later to deploy it (pip install -U sequence-ai); its client runs on 0.56.0 or later with its video encoder (pip install -U "sequence-ai[observe]"), beside LeRobot 0.6.1 or later.

    • v5's files; only policy.py's comment on CPU and memory changed: the machine its card sets (an L4's 8 cores and 32 GiB), or what it declares where that is more.
    Use v6
    molmoact2-so101 v6shell
    # start a policy from this version, then deploy it
    # my_policy_user: the name you choose
    seq init molmoact2-so101@v6 my_policy_user
    seq deploy my_policy_user
    
    # or move a deployment made from this template to it
    seq policy upgrade my_policy_user --to molmoact2-so101@v6
    
    # on the robot's machine: this version's client.py and its README
    pip install -U sequence-ai
    seq init molmoact2-so101@v6 my_policy_user

    In the console, Deploy deploys this version: the console deploys the stable one.

  2. v5

    As v4; the README says a worker's CPU and memory are billed as the machine its card sets.

    Needs sequence-ai 0.56.0 or later to deploy it (pip install -U sequence-ai); its client runs on 0.56.0 or later with its video encoder (pip install -U "sequence-ai[observe]"), beside LeRobot 0.6.1 or later.

    • v4's files; only the README's billing of CPU and memory changed: the machine its card sets (an L4's 8 cores and 32 GiB), or what it declares where that is more.
    Use v5
    molmoact2-so101 v5shell
    # start a policy from this version, then deploy it
    # my_policy_user: the name you choose
    seq init molmoact2-so101@v5 my_policy_user
    seq deploy my_policy_user
    
    # or move a deployment made from this template to it
    seq policy upgrade my_policy_user --to molmoact2-so101@v5
    
    # on the robot's machine: this version's client.py and its README
    pip install -U sequence-ai
    seq init molmoact2-so101@v5 my_policy_user
  3. v4

    MolmoAct2 on an SO-100 or SO-101 as before; the README says which LeRobot to run it beside.

    Needs sequence-ai 0.56.0 or later to deploy it (pip install -U sequence-ai); its client runs on 0.56.0 or later with its video encoder (pip install -U "sequence-ai[observe]"), beside LeRobot 0.6.1 or later.

    • LeRobot 0.5 and earlier require a version of av that cannot be installed beside sequence-ai[observe]; 0.6.1 and later install with it.
    • LeRobot 0.7.0 reads the arm as 0.5 does, so from_lerobot() and to_lerobot() are unchanged.
    Use v4
    molmoact2-so101 v4shell
    # start a policy from this version, then deploy it
    # my_policy_user: the name you choose
    seq init molmoact2-so101@v4 my_policy_user
    seq deploy my_policy_user
    
    # or move a deployment made from this template to it
    seq policy upgrade my_policy_user --to molmoact2-so101@v4
    
    # on the robot's machine: this version's client.py and its README
    pip install -U sequence-ai
    seq init molmoact2-so101@v4 my_policy_user
  4. v3

    MolmoAct2 on an SO-100 or SO-101 as before; the README says which weights from_lerobot() and to_lerobot() are for.

    Needs sequence-ai 0.56.0 or later to deploy it (pip install -U sequence-ai); its client runs on 0.56.0 or later with its video encoder (pip install -U "sequence-ai[observe]").

    • from_lerobot() and to_lerobot() are for the official weights, trained before LeRobot's PR #777; a fine-tune on data LeRobot 0.5 recorded takes its readings and plays its rows as they are.
    • Sent raw to the official weights, elbow_flex's target sits 90 degrees off and shoulder_lift's is mirrored, about 290 degrees off at rest.
    • LeRobot 0.5 reads the arm with use_degrees=True, its default: the five joints in degrees, the gripper 0 to 100.
    Use v3
    molmoact2-so101 v3shell
    # start a policy from this version, then deploy it
    # my_policy_user: the name you choose
    seq init molmoact2-so101@v3 my_policy_user
    seq deploy my_policy_user
    
    # or move a deployment made from this template to it
    seq policy upgrade my_policy_user --to molmoact2-so101@v3
    
    # on the robot's machine: this version's client.py and its README
    pip install -U sequence-ai
    seq init molmoact2-so101@v3 my_policy_user
  5. v2

    MolmoAct2 on an SO-100 or SO-101 as before, with from_lerobot() and to_lerobot() for an arm LeRobot 0.5 reads.

    Needs sequence-ai 0.56.0 or later to deploy it (pip install -U sequence-ai); its client runs on 0.56.0 or later with its video encoder (pip install -U "sequence-ai[observe]").

    • Pass what LeRobot 0.5 reads through from_lerobot() and play each answer through to_lerobot(): sent raw, its targets sit 90 degrees off on two joints and the arm drives into its limit.
    Use v2
    molmoact2-so101 v2shell
    # start a policy from this version, then deploy it
    # my_policy_user: the name you choose
    seq init molmoact2-so101@v2 my_policy_user
    seq deploy my_policy_user
    
    # or move a deployment made from this template to it
    seq policy upgrade my_policy_user --to molmoact2-so101@v2
    
    # on the robot's machine: this version's client.py and its README
    pip install -U sequence-ai
    seq init molmoact2-so101@v2 my_policy_user
  6. v1

    Ai2's MolmoAct2 for an SO-100 or SO-101 arm, on an L4.

    Needs sequence-ai 0.56.0 or later to deploy it (pip install -U sequence-ai); its client runs on 0.56.0 or later with its video encoder (pip install -U "sequence-ai[observe]").

    Use v1
    molmoact2-so101 v1shell
    # start a policy from this version, then deploy it
    # my_policy_user: the name you choose
    seq init molmoact2-so101@v1 my_policy_user
    seq deploy my_policy_user
    
    # or move a deployment made from this template to it
    seq policy upgrade my_policy_user --to molmoact2-so101@v1
    
    # on the robot's machine: this version's client.py and its README
    pip install -U sequence-ai
    seq init molmoact2-so101@v1 my_policy_user