A robot policy can recognize what is in a scene yet still fail to reason about where things are in 3D. SLA tackles that gap before task-specific adaptation: it teaches the visual representations of a pretrained policy to align with spatial features across varied robot data, aiming to make geometry a transferable capability rather than a benchmark-specific add-on.
Key Findings
- Action prediction and spatial alignment are trained together, so geometric understanding develops alongside control rather than being bolted on only during downstream tuning.
- A frozen 3D perception teacher uses RGB and whatever depth or camera-calibration data is available. This lets the training corpus retain demonstrations with incomplete geometric annotations.
- Feature-dependent gates combine visual features from multiple policy layers, while state dropout discourages the policy from relying on memorized robot-state trajectories instead of visual cues.
- The paper reports state-of-the-art results on RoboDojo-Real, RoboTwin 2.0 Clean2Random, RoboCasa-Panda, and LIBERO-Plus, and a 24-percentage-point average success-rate improvement over π0.5 across eight real-world deployment conditions.
Method and Trade-offs
The spatial teacher, alignment projector, and layer gates are training-only components; inference keeps the original policy architecture and does not require depth maps or camera calibration as additional inputs. This avoids adding deployment-time computation, but shifts the extra modeling and data-processing burden to training. Results also depend on the quality and coverage of the demonstrations and the teacher’s spatial representations.
Who It’s For
Great fit for researchers developing general-purpose robot manipulation policies who want spatial knowledge to transfer across scenes, tasks, and embodiments. Look elsewhere if you need a standalone 3D perception model, or if your goal is to add spatial supervision without undertaking continual pre-training and downstream evaluation.
