The ability to accurately estimate distances from RGB image input is just at the frontier of current AI model capabilities.
Nonetheless, distance estimation is a critical for perception and planning in embodied AI applications like robotics which must navigate around our 3D world.
We just released SpaceThinker, a 3B open-weight VLM designed specifically for spatial reasoning tasks like distance and size estimation from RGB images. It’s small and fast enough for on-device use, trained entirely on open-source data/code.
Interesting finding: By switching model name in this colab, using the non-reasoning variant SpaceQwen (https://huggingface.co/remyxai/SpaceQwen2.5-VL-3B-Instruct), you'll find using the step-by-step reasoning prompt actually hurts performance, challenging the convention that reasoning models don't benefit from complex instructions the way non-reasoning models do.
Feedback, suggestions, and collaborators are welcome!
If you're interested in contributing, we open-sourced VQASynth—our implementation of the SpatialVLM approach for generating VQA-style datasets from images. VQASynth was used to create the SpaceThinker dataset, which powered the fine-tuning of the SpaceThinker model showcased here.
1 comment
[ 3.0 ms ] story [ 13.4 ms ] threadNonetheless, distance estimation is a critical for perception and planning in embodied AI applications like robotics which must navigate around our 3D world.
We just released SpaceThinker, a 3B open-weight VLM designed specifically for spatial reasoning tasks like distance and size estimation from RGB images. It’s small and fast enough for on-device use, trained entirely on open-source data/code.
* Model: https://huggingface.co/remyxai/SpaceThinker-Qwen2.5VL-3B
* Data: https://huggingface.co/datasets/remyxai/SpaceThinker
On the QSpatial++ benchmark, SpaceThinker sits between GPT-4o and Gemini 2.5 Pro in performance, see this comparison table: https://huggingface.co/remyxai/SpaceThinker-Qwen2.5VL-3B#qsp...
Interesting finding: By switching model name in this colab, using the non-reasoning variant SpaceQwen (https://huggingface.co/remyxai/SpaceQwen2.5-VL-3B-Instruct), you'll find using the step-by-step reasoning prompt actually hurts performance, challenging the convention that reasoning models don't benefit from complex instructions the way non-reasoning models do.
Additional resources:
* https://github.com/andrewliao11/Q-Spatial-Bench-code/blob/ma...
* https://huggingface.co/blog/NormalUhr/deepseek-r1-explained#...
Feedback, suggestions, and collaborators are welcome!
If you're interested in contributing, we open-sourced VQASynth—our implementation of the SpatialVLM approach for generating VQA-style datasets from images. VQASynth was used to create the SpaceThinker dataset, which powered the fine-tuning of the SpaceThinker model showcased here.