remyxai/SpaceOm
remyxai/SpaceOm is a 3 billion parameter vision-language model, fine-tuned from UCSC-VLAA/VLAA-Thinker-Qwen2.5VL-3B, with a 32768 token context length. It enhances spatial reasoning and robotics domain understanding by incorporating the o_proj module in LoRA fine-tuning and training on specialized datasets with longer reasoning traces. This model is particularly optimized for complex spatial understanding and visual question answering in robotics applications.
Loading preview...
SpaceOm: Enhanced Spatial Reasoning Vision-Language Model
SpaceOm is a 3 billion parameter vision-language model developed by remyxai, building upon the SpaceThinker model. It significantly improves spatial reasoning and understanding, particularly for robotics applications, through several key enhancements:
Key Capabilities & Enhancements
- Improved Reasoning: Fine-tuned with the
o_projmodule in LoRA, a technique shown to be crucial for reasoning models, and trained on datasets with longer reasoning traces to encourage more extensive 'thinking' processes. - Robotics Domain Alignment: Incorporates the Robo2VLM-Reasoning dataset, derived from Robo2VLM-1, to better align the model with robotics domain knowledge and Multi-Choice Visual Question Answering (MCVQA) examples.
- Synthetic Data Training: Utilizes synthetic reasoning traces generated from the SpaceThinker dataset and a 3D scene reconstruction pipeline (Molmo, VGGT, SAM2), inspired by SpatialVLM, to ground its spatial understanding.
Performance Highlights
SpaceOm demonstrates competitive performance on various spatial reasoning benchmarks:
- OmniSpatial: Achieves an overall score of 41.79, showing improvements over its predecessor, SpaceThinker-Qwen2.5-3B.
- SpatialScore: Scores 48.15 overall, with strong performance in Object-Location (55.24) and Positional-Relation (61.83) categories, and a notable 18.83 on the challenging SpatialScore-Hard split.
- SpaCE-10: Records an overall score of 32.32, excelling in Size Assessment (47.63) and Object-Object spatial relations (50.00).
- MindCube: Outperforms the Qwen2.5-VL-3B-Instruct base model, averaging 41.48% accuracy across tested cog map output settings, particularly in configurations involving augmented cognitive maps and feed-forward reasoning.
Good For
- Robotics Applications: Ideal for tasks requiring detailed spatial understanding and reasoning in robotic environments.
- Complex Visual Question Answering: Excels in scenarios demanding multi-step reasoning based on visual input.
- Research in Spatial AI: A valuable base for further research and development in spatial cognition for AI models.