[hujc7]

Review Map

• Exact changes: a stacked PR's page shows the cumulative diff of its dependency chain; the link pins the commit range that is the PR's own contribution.
• Links pin specific SHAs and can go stale after a branch update — this table on [DO-NOT-MERGE][Task Clean-up] Dexterous: lumped validation branch (series reference) #6324 is refreshed first.

PR	Status	Depends on	Exact changes
📌 #6421 Part 6/10: Handover + camera manager counterparts		—	changes
#6410 Part 7/10: Environment overview docs		#6421	changes
#6415 Part 9/10: Benchmark success-rate utilities		#6421	changes
#6582 Part 10/10: Warp variants → experimental (draft; merges last)		—	changes
#6324 [DO-NOT-MERGE] Lumped validation reference		ALL	—

Summary

Completes manager coverage of the dexterous task families by adding manager-based
counterparts for the Shadow handover and Shadow camera reorientation tasks.

• 2 manager tasks added: Isaac-Shadow-Handover and Isaac-Reorient-Cube-Shadow-Camera.
• The last 2 -Play registrations in isaaclab_tasks are removed.

1. Handover manager counterpart

The fused handover reward became a plain reward term, with success and goal-distance
bookkeeping moved to HandoverCommand, which owns the goal. It reports
Metrics/success_rate as a per-episode bit, matching the handover Direct environment.

HandoverCommand.command builds its pose per call rather than returning a persistent
buffer, so consumers that store it across steps are not aliased to live data.

The task keeps no observation, reset or event code of its own. Every observation is a
framework term or a reorientation one: fingertip pose is body_pose_w, the object block
is the root pose, root velocities, the generated command and the reorientation
goal-quaternion difference, and only the fingertip velocity stays task-local because the
framework has no body_vel_w. Both hands reset through the reorientation hand reset, one
term per hand. The object and goal block was previously observed once per agent with
identical values, so a single copy now carries the same information.

presets=randomized adds the domain-randomization terms to the reset, matching how the
reorientation manager tasks expose theirs.

The reset and action observations use the framework terms rather than task-local ones,
following Part 5: the object resets through reset_root_state_with_random_orientation
and both hands observe last_action, leaving only the two-hand joint and PD-target
reset task-local. This samples the object's reset orientation uniformly over SO(3)
where the Direct task composes X/Y rotations, and reports zero actions on the reset
step where the Direct task reports the pre-reset action.

2. Camera manager counterpart

The camera task runs on PhysX by default: the RTX render modalities require Fabric
cloning, which Newton does not support, so the inherited default could not render.
Newton remains selectable with physics=newton_mjwarp for the state-only observation
groups.

The feature-extractor observation term ignored its declared feature_extractor_cfg
parameter and read the environment configuration instead; it now honours the parameter.

3. Camera playback moves to play_mode

Isaac-Reorient-Cube-Shadow-Camera-Play and -Camera-Direct-Play were the last -Play
registrations left in isaaclab_tasks; every other task moved to play_mode in
#6676 — Remove the _PLAY configs. Their
playback environment count and inference feature extractor now live in play_mode
overrides on the manager and Direct camera configurations.

The override mutates the feature extractor rather than replacing it. Replacing the
object resets every field the caller does not name, which silently re-enabled the CNN
in ShadowHandCameraBenchmarkEnvCfg — a configuration whose entire purpose is to
disable it.

4. Reorientation action configuration

The action term is named through a module path, so loading a task configuration no
longer imports the USD bindings.

Validation

Trained at 4096 environments, 1500 iterations, seed 42, on one L40
(jichuanh-autorun-8gpu). Success rate is logged per reset batch, so the figures below
are the mean and range over the final eight batches rather than a single value.

Task	Backend	Success rate	Range
Isaac-Shadow-Handover (manager)	newton_mjwarp	0.756	0.745 - 0.779
Isaac-Shadow-Handover-Direct (reference)	newton_mjwarp	0.697	0.563 - 0.875
Isaac-Shadow-Handover (manager)	isaacsim_physx	0.999	0.992 - 1.000

The manager task matches the Direct reference it replaces on the same backend. The
spread between backends is a pre-existing property of the Newton Shadow Hand actuation,
not something this pull request introduces; it is tracked separately.

Isaac-Reorient-Cube-Shadow-Camera was smoke-tested rather than trained to convergence:
300 iterations at 256 environments through the Kit renderer completes cleanly. Its
rendering matrix and golden images are covered separately.

presets=randomized converges more slowly than the default and does not plateau within
1500 iterations. A longer run is in progress to set an appropriate iteration budget.

[greptile-apps]

Greptile Summary

This PR completes manager coverage of the dexterous task families by adding Isaac-Shadow-Handover (manager-based counterpart to the Direct MARL handover env) and Isaac-Reorient-Cube-Shadow-Camera (manager-based camera reorientation task), and removes the last two -Play registrations in isaaclab_tasks, replacing them with play_mode overrides.

• HandoverCommand owns success bookkeeping and goal-distance metrics, routing Metrics/success_rate to the unified TensorBoard path; the goal pose is constructed fresh each call via torch.cat to avoid consumer aliasing.
• BALL_CFG consolidates the old ObjectCfg(PresetCfg) into a single RigidObjectCfg for both the Direct and manager envs, removing the Newton-specific variant that previously excluded PhysX-only rigid-body properties (solver_position_iteration_count, sleep_threshold, physics_material). Newton backend will now receive these settings.
• HandoverEventPresetCfg neatly separates the state reset (ResetEventCfg) from domain randomisation (RandomizationEventCfg) via multiple inheritance, with presets=randomized as the opt-in path.

Confidence Score: 5/5

• Safe to merge; the changes are well-scoped manager env additions with a straightforward reward term and command term, backed by reported validation results.
• The new HandoverCommand, handover_goal_distance_reward, and HandoverManagerEnvCfg are clean and follow established patterns from the reorient counterparts. The consolidation of BALL_CFG removes an explicit Newton-specific variant, but this is a physics-property question rather than a control-flow defect. No crashes, incorrect state transitions, or broken contracts were found. The two observations noted are quality concerns for future maintainers, not blockers.
• The handover_env_cfg.py and handover_manager_env_cfg.py files share the unified BALL_CFG; if Newton physics turns out to reject or warn on the PhysX-specific properties, both files need a backend-specific ball config restored.

Important Files Changed

Filename	Overview
source/isaaclab_tasks/isaaclab_tasks/core/handover/mdp/commands.py	New HandoverCommand class implementing the goal-pose command term. The reset() method has a subtle ordering consideration when converting env_ids to slice(None), and the _resample_command uses only X/Y axis rotations (intentionally, to match the Direct task's goal sampling). Overall logic is sound.
source/isaaclab_tasks/isaaclab_tasks/core/handover/mdp/rewards.py	Adds handover_goal_distance_reward for the manager env. The 2× multiplier on handover_reward (which itself returns 2×) yields a 4× total; the docstring explains this mirrors the Direct env's per-hand sum. Clean addition, no issues found.
source/isaaclab_tasks/isaaclab_tasks/core/handover/handover_manager_env_cfg.py	New manager env cfg mirroring the Direct MARL single-agent adapter. Uses BALL_CFG (plain RigidObjectCfg, not a PresetCfg) which includes PhysX-only rigid_props and physics_material settings that the old ObjectCfg Newton variant explicitly excluded. The preset split for events (ResetEventCfg vs HandoverEventCfg) is well structured.
source/isaaclab_tasks/isaaclab_tasks/core/handover/handover_env_cfg.py	Simplified to introduce BALL_CFG as a shared singleton replacing the old ObjectCfg(PresetCfg). The EventCfg class was removed (it was already noted as unwired). Changes are clean doc/cleanup, but the BALL_CFG consolidation carries the same Newton-backend concern as the manager file.
source/isaaclab_tasks/isaaclab_tasks/core/handover/init.py	Adds the Isaac-Shadow-Handover manager env registration before the Direct registration. Correctly wires env_cfg_entry_point and rsl_rl_cfg_entry_point. No issues.
source/isaaclab_tasks/isaaclab_tasks/core/handover/mdp/init.pyi	Stub file updated to export HandoverCommand, HandoverCommandCfg, and handover_goal_distance_reward. Correct additions, consistent with the new modules.

Sequence Diagram

sequenceDiagram
    participant CM as CommandManager
    participant HC as HandoverCommand
    participant CT as CommandTerm (base)
    participant EER as EpisodeErrorRecorder

    Note over CM,EER: Per-step flow
    CM->>HC: compute(dt)
    HC->>CT: (base) _update_metrics()
    HC->>CT: (base) _update_command() [no-op]

    Note over CM,EER: Episode reset flow
    CM->>HC: reset(env_ids)
    HC->>HC: "metrics["success_rate"][env_ids] = _succeeded.float()"
    HC->>CT: super().reset(env_ids)
    CT->>CT: mean metrics → extras dict
    CT->>CT: zero metrics
    CT->>HC: _resample_command(env_ids)
    HC->>HC: sample random X/Y-axis goal quat
    CT-->>HC: return extras
    HC->>HC: "_succeeded[env_ids] = False"
    HC->>EER: _minimum_goal_distance.reset(env_ids)
    EER-->>HC: "{mean, min, max} statistics"
    HC->>HC: route Metrics/success_rate → env.extras["log"]
    HC-->>CM: return remaining extras

Loading

_{Reviews (2): Last reviewed commit: "Use the shared actuated-joint list in th..." | Re-trigger Greptile}

[greptile-apps]

Mixed tensor/scalar types in extras log dict. dist_reward and dist_goal are stored as zero-dim PyTorch tensors (.mean() without .item()), while the Metrics/* keys on the very next lines use .item() to produce Python floats. A logging backend that serializes extras["log"] to JSON or passes values to a scalar summarizer will silently fail or produce unexpected output for those two keys.

Suggested change

	env.extras.setdefault("log", {})["dist_reward"] = per_agent_reward.mean()
	env.extras["log"]["dist_goal"] = goal_distance.mean()
	env.extras["log"]["Metrics/goal_distance"] = goal_distance.mean().item()
	env.extras.setdefault("log", {})["dist_reward"] = per_agent_reward.mean().item()
	env.extras["log"]["dist_goal"] = goal_distance.mean().item()
	env.extras["log"]["Metrics/goal_distance"] = goal_distance.mean().item()

[hujc7]

Resolved. The reward term no longer writes to the extras log dict at all; the success metric is routed through the command term's metrics and namespaced as Metrics/success_rate by CommandManager.reset(). No mixed tensor/scalar values remain.

[ooctipus]

Do not put cfg together in implementation

[hujc7]

Will be addressed in a followup PR for asset consolidation.

[ooctipus]

something is wrong in this cfg, we shouldn't work around the asset and behavior different like this.....

[ooctipus]

this looks like a very inefficient way to sample in quaternion space, you can just randomly sample r p y, then convert to quaternion instead...

[ooctipus]

.............. can we fix the issue from root not patch like that