Found first: a primary source the press has not covered yet.
A research team has posted a technical report on a task-agnostic agent harness for zero-shot robotic manipulation that ranks first on the RoboDojo leaderboard with a 57.7% success rate, more than doubling the performance of the same vision-language model under RoboDojo's reference control interface. The Auriga technical report, submitted October 31 2026, closes 68.3% of the planning-execution gap separating current VLM-based robot policies from their arm-free upper bound.
What the source says
Auriga wraps a VLM in an explicit task belief, calibrated 3D measurements, dual-arm control via direct tool calls or multi-step Python programs, visual trajectory previews with paused simulation, and runtime monitoring of stage tolerances and grasp bounds. The planning-execution gap the paper measures runs from 22.5% (GPT-6 Astra under the reference control interface) to 76.2% (the arm-free upper bound); Auriga closes 68.3% of that distance, reaching a 64.0 score and 57.7% success rate on RoboDojo. On RoboCasa365 the system reaches 73.2%, outperforming VLA-assisted agents by up to 23.1% on composite tasks. Under scene randomization, Auriga remains first-ranked while learned policies drop by up to 33.9%. The system requires zero target-task exploration, demonstrations, or environment resets, and went through 13 cross-task capability iterations before evaluation was frozen.
Why it matters
The result puts harness engineering alongside model selection and policy training as a first-class variable on manipulation benchmarks. Wrapping the same VLM in structured belief, calibrated sensing, and a code execution gateway produces a larger gain than the entire gap between reference-interface VLMs and VLA-assisted agents. The scene-randomization result adds weight to that claim: a drop of up to 33.9 percentage points for learned policies with no corresponding drop for Auriga points to a generalization advantage that gradient-trained approaches are not capturing. For practitioners building robot agents, the control interface deserves as much scrutiny as the model behind it.