Repository navigation
Conversation
github-actions
Bot
added
documentation
This was referenced Aug 7, 2026
Merged
[Task Clean-up][Manager] Dexterous Part 6/10: Add the handover and camera manager counterparts
#6421
Merged
Adds manager-based Allegro and Shadow cube reorientation environments that mirror their Direct counterparts, sharing the reward, termination, command and event definitions so the two workflows stay numerically in step. A value-parity test pins the shared task-defining values. Moves the shared task helpers into the reorientation task package and renames the manager Allegro configurations after the robot rather than the object.
Adds the manager-based handover task and the Shadow camera reorientation task, each mirroring its Direct counterpart's observation, reward and termination contracts. Extends the value-parity test to the handover pair. Retargets the handover recorder import to the reorientation helpers, which the previous part moved out of core/utils.py.
The four OpenAI identifiers reproduced a specific sim-to-real setup -- 20 Hz control, action and observation noise, and an episode budget spent per goal -- which does not generalize to other reorientation work. They now live in contrib as two tasks, with the feed-forward and recurrent policies sharing one environment and selected with --agent. The observation architecture does generalize, so it stays in the core task as presets=openai: a 42-dimensional actor restricted to what a physical hand can measure, paired with a 187-dimensional privileged critic. One preset token switches both the environment and the RSL-RL agent. Replaces the obs_type string with the independent reduced_obs and asymmetric_obs flags.
ooctipus
reviewed
Aug 7, 2026
| return SceneEntityCfg(name, joint_names=".*") | ||
|
|
||
|
|
||
| def _fingertip_entity(name: str) -> SceneEntityCfg: |
Collaborator
There was a problem hiding this comment.
you can set a constant instead of create a funciton
ooctipus
reviewed
Aug 7, 2026
| ) | ||
|
|
||
|
|
||
| def _hand_entity(name: str) -> SceneEntityCfg: |
Collaborator
There was a problem hiding this comment.
constant instead of function
ooctipus
reviewed
Aug 7, 2026
| # soft limits equal the hard limits here: soft_joint_pos_limits_factor defaults to 1.0 | ||
| right_joint_pos = ObsTerm(func=mdp.joint_pos_limit_normalized, params={"asset_cfg": _hand_entity("right_hand")}) | ||
| right_joint_vel = ObsTerm(func=mdp.joint_vel, scale=0.2, params={"asset_cfg": _hand_entity("right_hand")}) | ||
| right_fingertip_pos = ObsTerm(func=mdp.fingertip_pos, params={"asset_cfg": _fingertip_entity("right_hand")}) |
Collaborator
There was a problem hiding this comment.
why not use the one function to return finger tip state?
hujc7
force-pushed
the
jichuanh/reorient-openai-contrib-p8
branch
from
August 7, 2026 10:05
75eee1e to
b0d2d84
Compare
ooctipus
reviewed
Aug 7, 2026
| rewards: OpenAIRewardsCfg = OpenAIRewardsCfg() | ||
| terminations: OpenAITerminationsCfg = OpenAITerminationsCfg() | ||
|
|
||
| enable_domain_randomization: bool = True |
Collaborator
There was a problem hiding this comment.
does this needs to exist?
ooctipus
reviewed
Aug 7, 2026
| physics=PhysicsCfg(), | ||
| ) | ||
| # a looser goal, chased repeatedly, with falls punished | ||
| fall_penalty = -50.0 |
Collaborator
There was a problem hiding this comment.
we shouldn't need this kind of direc env style field creation/overrride if we are in manager style.
Collaborator
Author
|
fold into #6418. comment should be addressed there. |
hujc7
added a commit
that referenced
this pull request
Aug 11, 2026
…anager counterparts (#6418) ## 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 #6324 is refreshed first. | PR | Status | Depends on | Exact changes | |---|---|---|---| | #6411 Part 1/10: Newton cloner/cubric/visualizer fixes |  | — | merged | | #6412 Part 2/10: OVPhysX articulation + manager runtime |  | — | merged | | #6413 Part 3/10: Reorient Direct, torch |  | — | [changes](https://github.com/isaac-sim/IsaacLab/pull/6413/changes/f4895f0f9ee..d29afc75e71) | | #6414 Part 4/10: MARL-to-single-agent fix + handover/camera Direct |  | #6413 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6414/changes/d29afc75e71..b10a84948f8) | | 📌 #6418 Part 5/10: Reorient manager counterparts |  | #6413 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6418/changes/3ea6f7bbf6c..0271783ea17) | | #6421 Part 6/10: Handover + camera manager counterparts |  | #6413, #6414, #6418 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6421/changes/b31b04eaf71..0cae3d87be4) | | #6410 Part 7/10: Environment overview docs |  | #6421 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6410/changes/b50ac8906fc..af259c0778d) | | #6954 Part 8/10: OpenAI Shadow variants → contrib (draft) |  | #6421 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6954/changes/0cae3d87be4..b0d2d8448af) | | #6415 Part 9/10: Benchmark success-rate utilities |  | #6421 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6415/changes/b50ac8906fc..c7f2f019d8b) | | #6582 Part 10/10: Warp variants → experimental (draft; merges last) |  | #6413 | [changes](https://github.com/isaac-sim/IsaacLab/pull/6582/changes/d29afc75e71..167c28578b3) | | #6324 [DO-NOT-MERGE] Lumped validation reference |  | ALL | — | ## Summary Adds manager-based counterparts for the Shadow cube reorientation task and aligns the existing Allegro manager task with the Direct contracts. - **3 manager tasks** mirror their Direct counterparts: `Isaac-Reorient-Cube-Shadow`, `-Shadow-Camera` and `-Allegro`. - The OpenAI variants move to the contributed tasks **unchanged**. They encode a paper-specific training regime, so they get no manager counterpart. - `Metrics/success_rate` now means the same thing in both workflows. - The in-hand cube is one object per hand instead of one per physics backend. ## 1. Manager tasks match the Direct contracts Observations, actions, rewards, terminations, reset distributions and timing were brought to the Direct values. A value-parity test asserts the decimation / episode length / simulation step triple, the orientation success tolerance and the fall distance, so drift on either side fails CI. The Allegro observation space changes size, so existing manager checkpoints must be retrained. Domain randomization becomes opt-in through `presets=randomized`, replacing a configuration-file flag that the command line could not override. ## 2. OpenAI variants move to the contributed tasks `Isaac-Reorient-Cube-Shadow-OpenAI-{FF,LSTM}-Direct` become `IsaacContrib-Reorient-Cube-Shadow-OpenAI-{FF,LSTM}-Direct`. Both identifiers are kept: they share one environment configuration and differ only in their agent configuration, so collapsing them would leave the recurrent policy unreachable — the training entry point selects an agent configuration by workflow, not by name. ## 3. One success rate, defined once `Metrics/success_rate` named two quantities. The manager tasks reported goals reached over goals presented; the Direct tasks reported a per-episode success bit drawn at `success_count_threshold`. The curves were not comparable despite sharing a name and measuring the same task. Both now report the per-attempt rate, and `success_count_threshold` is removed as unused. The goal accounting moves into a `SuccessTracker` beside `EpisodeErrorRecorder`, which also suppresses success detection for the single step after an automatic reset — where the goal has been drawn but no physics has run against it, so a reset pose inside the success threshold would otherwise bank a goal the policy never earned. ## 4. One in-hand cube per hand, not per backend The cube was a `PresetCfg` with a variant per physics backend, so each hand manipulated a different object depending on the backend. On Shadow the branches disagreed on size and spawn height: PhysX spawned a 60 mm rigid body at `z=0.6`, Newton a 54 mm **articulation with no joints** at `z=0.535`. A policy trained on one backend was not solving the same task as one trained on the other. Each hand now declares a single `CUBE_CFG`. Shadow keeps the PhysX values; Allegro keeps its 1.2 scale. The `mass_props` density overrides are dropped: the asset authors an absolute `physics:mass` of 0.216 kg, which takes precedence over density, so they never applied on either backend. ## 5. Configuration hierarchy flattened, classes named by role Every manager task overrode all seven of the shared base's sub-configurations, so the base carried pre-alignment defaults that each task then undid. The environment configurations now derive from `ManagerBasedRLEnvCfg` directly. Removed: `ReorientObjectEnvCfg`, the shared sub-configurations no task constructed, and `reorient_common`. `isaaclab_tasks.core.utils` moves to `isaaclab_tasks.core.reorient.utils`, which the hand-over task already imports from. One naming convention now applies throughout, since `SceneCfg`, `EventCfg` and `ObservationsCfg` are each defined in dozens of files and gave no clue which base a task was specializing. Classes shared across files carry the task prefix; the `Base` suffix marks only those a subclass must complete by supplying fields the parent leaves unset, so the observation halves are named as parts rather than templates. Classes nested inside another configuration drop the prefix. Registered and preset-selectable configurations always name their task and hold fields only, with their helpers declared beside them. ## Validation Eight arms, 1500 iterations, 8192 environments, `physics=newton_mjwarp`, one L40 each. All exited cleanly with no regression in reward. | Arm | reward before | reward after | |---|---|---| | `Isaac-Reorient-Cube-Shadow` | 67.08 | 66.24 | | `-Shadow` `presets=asymmetric` | 42.08 | 52.95 | | `-Shadow-Direct` | 3532.95 | 3967.58 | | `-Allegro` | 25.22 | 34.03 | | `-Allegro` `presets=randomized` | 16.03 | 39.59 | | `-Allegro-Direct` | 656.78 | 1003.11 | | `IsaacContrib-…-OpenAI-FF-Direct` | 2757.56 | 4533.21 | | `IsaacContrib-…-OpenAI-LSTM-Direct` | — | 2585.54 | The LSTM arm resolves `ShadowHandAsymLSTMPPORunnerCfg`, confirming the recurrent policy is reachable through its own identifier. ### Known failure `test_rendering_shadow_hand[newton-isaacsim_rtx-rgb]` fails; it passes on `develop`. Only the `rgb` modality is affected — depth, normals and all three segmentations pass on the same backend, so the cube move is not an obvious explanation. Root cause is not yet established.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Review Map
Summary
--agent.presets=openai— a 42-dimensional actor limited to what a physical hand can measure, plus a 187-dimensional privileged critic. One preset token switches both the environment and the RSL-RL agent.obs_typestring with independentreduced_obs/asymmetric_obsflags, so a task may narrow the actor, add a privileged critic, or both.presets=openaion the manager workflow) added rather than lost.Task identifiers
Isaac-Reorient-Cube-Shadow-OpenAI-FFIsaacContrib-Reorient-Cube-Shadow-OpenAIIsaac-Reorient-Cube-Shadow-OpenAI-LSTMIsaacContrib-Reorient-Cube-Shadow-OpenAI --agent rsl_rl_lstm_cfg_entry_pointIsaac-Reorient-Cube-Shadow-OpenAI-FF-DirectIsaacContrib-Reorient-Cube-Shadow-OpenAI-DirectIsaac-Reorient-Cube-Shadow-OpenAI-LSTM-DirectIsaacContrib-Reorient-Cube-Shadow-OpenAI-Direct --agent rsl_rl_lstm_cfg_entry_pointOn the removal without deprecation aliases
AGENTS.mdrequires prior deprecation for public API removals. These four identifiers are being removed outright, deliberately: the equivalent tasks exist underIsaacContrib-names in the same release, the changelog fragment carries the full mapping above, and retaining four alias registrations plus anobs_typetranslation would preserve a naming scheme this series exists to retire. Raised in review — happy to add aliases if maintainers prefer.Series
Part 8 of 10. Depends on Part 6/10 (#6421), which owns
Isaac-Reorient-Cube-Shadow-Camera; this branch is based on it, so the diff contains #6421's commits until it merges. Review from75eee1e7211onward.Validation
test_reorient_value_parity.py,test_reorient_utils.py,test_hydra.py)presets=openaiverified on the Direct and manager workflowsupdate_environments_rst.py --checkregenerated