Robotics researcher: harness + tool calls beat raw VLA on control benchmarks, at 2.3x lower cost

rbhar90 · x · 2026-09-08

A Berkeley/NVIDIA robotics researcher argues that harnesses plus tool calls (IK, SAM3, etc.) benchmark multimodal models for continuous robot control far more cheaply, reliably and quickly than end-to-end control — one task jumped from 40% to 95% with 6.2x fewer output tokens at 2.3x lower cost. Self-evolving harnesses and skill libraries will rapidly expand the capability frontier, with VLA and WAM as parts of those libraries. Robotics and physical-world applications are the next frontier for multimodal models and agents.

Original post →

More from Embodied

Embodied channel →