Research note · October 24, 2026

What Can a Decision Model Do for Robot Manipulation?

A 4B model (Qwen3.5-4B + LoRA, pointer readout) · 184,942 training questions · public train splits only · 8 comparison models in one harness