Skip to content

decision: expose joint acceleration for source-compatible task metrics #1543

Description

@TATP-233

Problem

wuji-mjlab logs Episode_Metrics/joint_acceleration_rms, but a downstream task consuming UniLab's public Entity API cannot obtain joint acceleration from the mjwarp backend. During wuji_unilab metric-name alignment, all other source metrics could be mapped without backend-private access.

Evidence

  • The source metric reads robot.data.joint_acc and computes RMS across selected joints.
  • UniLab EntityData exposes joint position and velocity but no joint acceleration.
  • UniSim has no declared qacc/joint-acceleration state field.
  • A finite difference of public joint velocity would be a different, control-step quantity and must not be silently logged under the source native-acceleration name.

Requested maintainer decision

Choose whether to add a declared backend-neutral joint-acceleration state contract, expose an explicitly named control-step finite-difference metric at the task layer, or keep this source metric unsupported. The Wuji downstream currently chooses the last option and reports the omission.

Scope

This does not block Wuji training or the other metric-name alignments. Any new public backend contract requires owner design, tests, and adapter behavior before downstream consumption. No fallback is requested.

Related: unilabsim/wuji_unilab#17 and roadmap unilabsim/wuji_unilab#1.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions