Repository navigation
GenManip v1.1.0 brings simulation and evaluation updates for Wet-OP Bench, with new scientific experiment metrics, reliable evaluation requests, and improved demonstration recording and rendering.
Highlights
Scientific Experiment Metrics
Add evaluation metrics for laboratory manipulation tasks, including liquid pouring, stirring, funnel use, lid removal, centrifuge tube manipulation, weighing, titration, and chemical reactions.
Reliable Evaluation Requests
Introduce a request protocol with deduplication, pending responses, and cached result replay. Clients using this protocol can retry requests without executing completed actions again.
Request tracing, error reporting, and worker recovery improvements also make evaluation failures easier to diagnose.
Motion and Placement
- Add object-position anchors for custom motions, allowing targets to follow a point on an object while retaining world-frame orientation.
- Support local placement jitter and configurable rotation fallback.
- Fix placement validation to account for equivalent quaternion representations.
Scene Recovery and Rendering
- Improve restoration of particle states, material attributes, and articulation root poses.
- Extend demonstration recording and rendering support for particle-based liquids and material changes.
- Add configurable material color overrides.
Other Improvements
- Fix range and stacking metrics.
- Improve collider and mesh transform handling.
- Support custom USD asset paths for the Lift2 robot.
Release Scope
This release updates the GenManip simulation and evaluation code. Wet-OP Bench task configurations, runtime assets, and client submodule updates are not bundled in this change.
The reliable-request protocol requires a compatible client. Legacy evaluation endpoints remain available.
Related PR: #25