Appendix B.3 — Visual task examples
48 full-size case figures · 20 task families · 3 benchmark levels
These are the complete figures prepared for Appendix B.3. Each task page groups its input-modality variants. Click a figure to open its full-size PNG, or use the original PDF link for zooming and download.
The cards explain benchmark inputs and motion assets; they are not model-generated predictions. For Level 2, read the target/reference wording printed in each figure: some tasks show an original reference motion rather than a computed modified result.
Back to the RoboSteer dataset · Task media examples
Level 1
Full conditioning
- Text-to-Motion Generation — 1 figure
- Video-to-Motion Imitation — 1 figure
- Audio-to-Motion Generation — 1 figure
- Rhythm-to-Motion Alignment — 1 figure
- Rotation-to-Pose Generation — 1 figure
Temporal completion
- Motion Prediction — 5 figures
- Motion Retrodiction — 4 figures
- Motion Interpolation — 4 figures
- Key-frame Conditioning — 4 figures
Spatial completion
- Upper-to-Full Body Completion — 4 figures
- Lower-to-Full Body Completion — 4 figures
- Target Reaching — 2 figures
Level 2
Constraint steering
- Order — 3 figures
- Times — 2 figures
- Trajectory — 2 figures
- Amplitude — 2 figures
- Body Restrain — 2 figures
- Direction — 2 figures
- Speed — 2 figures
Level 3
Compositional steering
- Interleaved Multi-Source Steering — 1 figure