Spatial-TRM
A compact recursive reasoning model for spatial context prediction that combines a frozen vision-language feature extractor, a convex region decoder and a weight-shared recursive reasoning module, improving spatial reasoning accuracy through iterative refinement.
Timeline (last 90 days)
The Spatial-TRM model has only 11.2 million trainable parameters.
On the RoboSpatial-Home benchmark, Spatial-TRM achieves a context accuracy of 29.51%.
Compared with single-step models of the same parameter count, Spatial-TRM's context accuracy improves by 44.0% relatively and 9.02 percentage points absolutely.
Spatial-TRM has a total parameter count of 877.8 million (including the frozen vision-language feature extractor).
The RoboPoint model has 13 billion parameters, and Spatial-TRM's total parameter count is about 1/15 of that.
Spatial-TRM's architecture contains three key components: a frozen vision-language feature extractor, a convex region decoder, and a weight-sharing recursive reasoning module.
Spatial-TRM's recursive reasoning module supports six spatial relations, three reference frames, and open-vocabulary anchor objects.
The comparative experiment for Spatial-TRM is designed so that the two models have exactly the same parameter count, with the only difference being whether recursive refinement is used.
All Facts (8)
The Spatial-TRM model has only 11.2 million trainable parameters.
50%UnverifiedOn the RoboSpatial-Home benchmark, Spatial-TRM achieves a context accuracy of 29.51%.
50%UnverifiedCompared with single-step models of the same parameter count, Spatial-TRM's context accuracy improves by 44.0% relatively and 9.02 percentage points absolutely.
50%UnverifiedSpatial-TRM has a total parameter count of 877.8 million (including the frozen vision-language feature extractor).
50%UnverifiedThe RoboPoint model has 13 billion parameters, and Spatial-TRM's total parameter count is about 1/15 of that.
50%UnverifiedSpatial-TRM's architecture contains three key components: a frozen vision-language feature extractor, a convex region decoder, and a weight-sharing recursive reasoning module.
50%UnverifiedSpatial-TRM's recursive reasoning module supports six spatial relations, three reference frames, and open-vocabulary anchor objects.
50%UnverifiedThe comparative experiment for Spatial-TRM is designed so that the two models have exactly the same parameter count, with the only difference being whether recursive refinement is used.
50%