SpatialClaw is a training-free framework from NVIDIA Research designed to enhance spatial reasoning in vision-language models. By treating code as an action interface, the agent can compose perception tools and revise its logic through a stateful Python kernel rather than following rigid tool schemas or pre-set plans. This allows for complex geometric computations across multiple views or video frames without model retraining.
Key features:
- Composes perception outputs to allow iterative reasoning and error correction.
- Integrates depth estimation and segmentation tools via code execution.
- Follows a five-stage loop of planning, generation, execution, feedback, and submission.
- Shows significant performance improvements in multi-view and dynamic video tasks.
Key features:
- Composes perception outputs to allow iterative reasoning and error correction.
- Integrates depth estimation and segmentation tools via code execution.
- Follows a five-stage loop of planning, generation, execution, feedback, and submission.
- Shows significant performance improvements in multi-view and dynamic video tasks.
