Draft tracker: split Franka core migration

[maxkra15]

Superseded by four focused draft PRs against develop:

1. Remove unsupported Franka Reorient task #8168 — Remove unsupported Franka Reorient.
2. Make flat Franka the main config and fix backend body resolution #8169 — Add the opt-in flat Franka asset config and backend body resolution.
3. Migrate Franka tasks to the shared asset and add minimal presets #8170 — Migrate Reach, Drawer, and relevant contributed tasks (depends on Make flat Franka the main config and fix backend body resolution #8169).
4. Improve Franka Lift resets and continuous Reach tracking #8171 — Migrate rigid/deformable Lift and reset/ADR fixes (depends on Remove unsupported Franka Reorient task #8168 and Make flat Franka the main config and fix backend body resolution #8169).

Each PR selects the active-release backport checkbox. Review and merge prerequisites first; then rebase the dependent PRs. The public-S3 flat asset and checkpoint qualification remain separate merge gates. Camera checkpoints are out of scope.

Do not merge this original combined diff.

[maxkra15]

run-ci

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no new findings were introduced since the previous review, and all previous findings are manually resolved.

Summary

Restores the core Franka task contracts around a shared flat Menagerie asset.

• Makes the flat multi-physics Franka asset canonical while retaining an explicit legacy configuration.
• Selects backend-specific physics payloads and configurable collision variants across PhysX and Newton.
• Aligns actuator, mimic-finger, frame-transformer, controller, reset-clearance, and solver behavior.
• Adds affine differential-IK action offsets and Reorient training bootstrap behavior.
• Updates affected core tasks, contrib configurations, demos, tests, and changelog fragments.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Shared[Canonical flat Franka asset] --> Variants{Physics payload}
  Variants -->|physx| PhysX[PhysX backend]
  Variants -->|mujoco| Newton[Newton backend]
  Shared --> Colliders{Collider preset}
  Colliders --> Gripper[gripper_only default]
  Colliders --> Arm[primitive or convex arm contacts]
  PhysX --> Tasks[Reach / Lift / Reorient / Cabinet]
  Newton --> Tasks
  Shared --> Frames[Frame-transformer resolution]
  Frames --> Tasks
  Shared --> Actuators[Unified arm and mimic-finger contract]
  Actuators --> Tasks

Loading

_{Reviews (3) · Last reviewed commit: "Bootstrap Franka Reorient training"}

[isaaclab-review-bot]

Isaac Lab Review Bot

The canonical Franka asset migration is coherent across backend variants, task presets, DiffIK affine actions, and nested-body frame resolution, but it introduces one likely runtime failure in Lift/Reorient termination handling and two public-configuration compatibility issues that should be resolved before merge.

• Design and architecture: Using one flat Franka asset with explicit Physics and Colliders variants is consistent across the core tasks. However, deriving the unrelated Robotiq configuration from the new canonical config silently changes its inherited self-collision setting from enabled to disabled; that configuration should explicitly preserve its prior contract or derive from the legacy base.
• API: The new DifferentialInverseKinematicsActionCfg.offset is backward compatible because its zero default preserves existing action processing. In contrast, FRANKA_PANDA_CFG changes its actuator dictionary keys from panda_shoulder/panda_forearm to panda_arm and adds a passive-finger group. Existing consumers indexing the old public keys will fail, so the assets changelog needs to mark the compatibility change clearly and provide concrete migration guidance.
• Implementation: The frame-transformer filtering, collider selection, and OSC actuator-property propagation align with their consumers. The Lift out_of_bound term now passes root_pos_w, root_quat_w, and root_vel_w wrappers directly to torch.isfinite and tensor comparisons, unlike the prior and neighboring code that unwraps them through .torch; all three reads should use the tensor accessor.

Significant concerns. Posted 3 actionable findings inline.

Automated review; human maintainers own approval decisions.

[maxkra15]

run-ci

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[maxkra15]

run-ci

[maxkra15]

run-ci

[maxkra15]

@isaaclab-review-bot review

[maxkra15]

run-ci

[maxkra15]

run-ci

[maxkra15]

Superseded by #8168, #8169, #8170, and #8171. Each targets develop with active-release backport selected.