Guava open-sources VLM distillation: 2K sim trajectories yield a 4B robot agent at 90% real-world success
furongh · x · 2026-10-06
Researchers from the University of Maryland and others open-sourced Guava, a framework that distills frontier vision-language models into compact, locally deployable robot agents using purely simulation-generated interaction data.
- Approach: teacher (GPT-5.4) and student share the same observation/tool/feedback interface via a manipulation harness, so teacher interaction histories serve directly as supervision for the student's ReAct-style single-action replanning loop
- Training: 2,000 simulation trajectories are enough to fine-tune Qwen3.5-4B into Guava-4B
- Results: 90.0% real-world success (vs. 93.3% for the teacher), 7.1× lower model-call latency with local inference, generalization to unseen tasks and object configurations, and closed-loop recovery from execution failures
Paper, code, and project page are all public.
More from Embodied
- NormalFactory unveils FrontierCAD, a verifier for grading AI-generated CAD parts and assemblies — hudzah · 2026-10-06
- Walden Robotics CTO to discuss Physical AI for line troubleshooting at Operations Calling — adnothing · 2026-10-06
- Robot demo with "don't try this at home" warning goes viral — justin_hart · 2026-10-06
- 67-gram Stembot robot climbs plant stems and builds 3D models to autonomously hunt pests — LexiLove · 2026-10-06
- Figure to premiere Hark Pro tomorrow at 9am, available right after — adcock_brett · 2026-10-06
- Alexander Wolf is building an artificial brain, one neuron at a time — Scobleizer · 2026-10-06