Guava distills GPT-5.4 into a 4B robot agent using 2,000 sim trajectories, zero-shot on real hardware
ChongZzZhang · x · 2026-11-21
Guava at a glance
- Researchers from UMD and collaborators released Guava, an open-source pipeline that distills frontier VLM manipulation ability into a compact agent via a shared manipulation harness and simulation-generated data.
- Method: GPT-5.4 acts as teacher, collecting 2,000 trajectories under a unified observation/tool/feedback interface; those ReAct-style replanning histories directly supervise fine-tuning of Qwen3.5-4B into Guava-4B.
- Results: Guava-4B matches its teacher across simulation, real-world, and external benchmarks; 400x faster than Pear and 40x more token-efficient, with zero-shot transfer to real-world manipulation.
- The poster argues free sim data + VLM sim2real robustness + reasoning-driven generalization may remove the need for RL.
More from Embodied
- Facet-SLAM Hits 113 FPS on RTX 4090 with Triangle Splatting Real-Time SLAM — jonstephens85 · 2026-11-21
- NVIDIA Cosmos 3 hands-on: world model forgets the room after 8 seconds off-camera — PrateekVJoshi · 2026-11-21
- RoboCap-10k: 10,000 hours of open egocentric robot data across 11 countries — pabbeel · 2026-11-21
- Paralyzed man plays GTA VI using only his brain via Neuralink — mark_k · 2026-11-21
- ROS weekly: Colcon web dashboard, MOCAP4ROS 2.0, ROSCon Belgium — kscottz · 2026-11-21
- Neuralink patient plays GTA VI using only thoughts via brain implant — Polymarket · 2026-11-21