Integrate scene workspace checks and add a UR5 showcase

[yuecideng]

Description

Prepare and restore B=1 Task Program scene variants through the Gym reset lifecycle, and expose robot-aware workspace checks as reusable sim.scene services. The UR5 round-table showcase makes the preparation decision visible: reject a red start before planning, then generate measured pick-and-place demonstrations from green starts.

• Keep RobotSceneWorkspace and SceneWorkspaceCandidate in sim.scene.workspace. Core checks return SceneCheckReport; the Task Program compatibility adapter returns ValidationResult.
• Retain host ownership, reset order, fixed-parameter drift checks, contact refresh, recording initialization and fresh runtime/planner bindings. Physical restoration requires a numeric round-trip check and unchanged ownership, topology, backend and control period.
• Fix static rigid-body property capture: static bodies have no dynamic mass/inertia data, while friction and damping remain checked.
• Keep all UR5 demo code under scripts/tutorials/scene/: robot_workspace.py owns the CLI and orchestration; four private sibling helpers own the scene/candidates, Task Program trial, display/capture and presentation video. Demo code is no longer shipped or documented as a public SDK API.
• Reuse real FK caches and one fixed downward grasp. Upright adjustments require fresh IK, current-limit and full-pose FK checks. The actual tabletop query results supply the red/green point cloud.
• Validate whole-cube support against the circular table, settling, measured task outcome, the first recorded cube pose and physical/controller restoration. A rejected start executes zero expert controls.

The task gate requires program completion, at least 5 cm of measured lift, and a released cube within 4 cm of the goal for ten consecutive control steps at linear speed <= 0.03 m/s and angular speed <= 0.2 rad/s. Accepted trials save trajectory evidence; LeRobot writing is disabled by default.

Dependencies: shared scene services from #761. Existing package dependencies cover this implementation. Refs #760 and #748. Scene proposal imports use the canonical sim.scene.expansion namespace from #805.

Showcase and documentation

Showcase guide · Original MP4

The guide embeds the full looping GIF (960 x 540, 8 fps, about 5 MiB), retains the original 1280 x 720, 25 fps MP4 download, and explains fixed-grasp color semantics, commands, acceptance and saved artifacts. Links from the Robot Workspace feature page and Simulation and Control tutorial index are retained.

The included seed-7 recording checks 1,190 object origins on a 2.4 m diameter round table: 703 pass and 487 fail. The red example produces zero actions; all three green starts complete measured pick-and-place trials. The committed GIF, video and screenshots are unchanged by the demo relocation.

Type of change

• New feature
• Enhancement
• Bug fix
• Documentation update

Validation

• 413 focused production scene/workspace and Task Program host/integration tests passed; the default slow case was qualified separately.
• 227 CI, mapped-context and documentation checks passed.
• The native scene-host recording/restoration test passed using the production API.
• Direct CLI smokes passed after relocation: scene --view none, task --view none, and task --view offscreen, all with seed 7 and three candidates. Task smokes were launched outside the repository to check path resolution. Both task runs succeeded 3/3; red rejection, first-frame cube poses and restoration passed. Candidate identities match across modes and views.
• Offscreen output includes the 1280 x 720, 25 fps, 686-frame explanation video. The four moved private helpers are byte-for-byte unchanged.
• Sphinx reference build passed with the existing repository documentation warnings; no removed SDK demo references remain. Black 26.3.1, whitespace and context checks passed; API coverage is 2602/2602 exports.

Demo-specific unit, presentation and native-mode tests were removed with the SDK demo modules. Automated tests retain production workspace, state and host coverage; no tests import the tutorial implementation. The tutorial is checked through syntax/help checks and direct CLI smokes.

Native qualification used a temporary coherent DexSim Python overlay over the installed binary because the local SDK installation mixed versions. The qualified physical scope remains fixed-topology B=1, single-arm UR5, Default backend/CPU physics.

Context review: revise the scene-services overview to point to the tutorial-owned implementation; other ownership and runtime guidance remains accurate.

Checklist

• Run black . with Black 26.3.1.
• Review and update affected documentation and agent context.
• Keep reusable services in production modules and demo code in tutorials.
• Validate production contracts and run the relocated CLI.
• Review dependencies; no additional runtime package is required.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues remain from this review.

Summary

Adds robot-aware scene checks, preparation through Gym reset, saved-state restoration, and a UR5 red/green workspace showcase.

• Robot workspace checks turn valid samples into checked object grasp poses.
• Scene variants prepare and restore through the Gym reset cycle.
• The UR5 demo shows red starts stopping and green starts reaching pick-and-place.
• The API docs use the canonical scene expansion namespace.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Scene candidate] --> B[Gym reset and reset events]
  B --> C[Apply poses and settle]
  C --> D[Refresh contacts and check scene]
  D --> E{Checks pass?}
  E -->|No| F[Reject without expert actions]
  E -->|Yes| G[Seed recording and create fresh bridge]
  G --> H[Execute task and measure outcome]
  H --> I[Restore saved state and compare]

Loading

_{Reviews (13) · Last reviewed commit: "refactor(demo): move workspace showcase ..." · Reviewed by Greptile}