Guava distills GPT-5.4 into a 4B robot agent with just 2,000 simulated trajectories
jbhuang0604 · x · 2026-11-09
- Guava transfers frontier VLM manipulation capabilities to a compact, locally deployable agent via a shared manipulation harness (unified observation/tool/feedback interface) plus simulation-generated interaction data.
- Using GPT-5.4 as teacher, the team collected 2,000 trajectories and fine-tuned Qwen3.5-4B into Guava-4B.
- The compact agent approaches its frontier teacher across simulation, real-world, and external benchmarks while cutting inference latency and token usage substantially.
- The team ablated interface choices, including ReAct-style single-action replanning vs. multi-action program generation.
Collaboration across UMD, UIUC, Waterloo, MBZUAI, UPenn, and Amazon FAR; code open-sourced. A companion work adds active visual and physical feedback (View/Probe) for aerial manipulation coding agents.
More from Embodied
- Apple's HALO humanoid trains in 1000 simulated homes, hits 91% success in real houses — heyshrutimishra · 2026-11-10
- Boston Dynamics puts wheels on Atlas to stress-test its whole-body RL controller — CyberRobooo · 2026-11-09
- Dex-One2Many learns dexterous manipulation from one human video, 95%+ success across 4 hands — jbhuang0604 · 2026-11-09
- Fly-by-Code: coding agents pilot drones via active visual and physical feedback loops — jbhuang0604 · 2026-11-09
- μ0 world model learns robot physics from video only, beats π0 by 5 points on RoboCasa365 — jbhuang0604 · 2026-11-09
- NVIDIA unveils GR00T 2 open-world robot action model, hitting 95% success with 5 demos — CyberRobooo · 2026-11-09