Found first: a primary source the press has not covered yet.
A technical report submitted October 31, 2026, introduces Auriga, a task-agnostic agent harness for zero-shot robotic manipulation that takes first place on the RoboDojo leaderboard with a 57.7% success rate (64.0 score), more than doubling the same VLM running under the benchmark's reference control interface. The paper, by Chengcheng Wang and 20 co-authors, requires no target-task trials, demonstrations, or environment resets.
What the source says
Auriga wraps GPT-6 Astra with an explicit task belief, tool calls, synthesized Python programs, calibrated 3D measurements, dual-arm control, and runtime safety checks. The reference harness running the same model achieves 22.5% on RoboDojo; Auriga reaches 57.7%, closing 68.3% of the gap toward the 76.2% arm-free execution upper bound. On RoboCasa365 the system reaches 73.2%, outperforming VLA-assisted agents by up to 23.1% on composite tasks. Under scene randomization, Auriga holds its leaderboard position while learned policies degrade by up to 33.9%.
Why it matters
The same underlying model goes from 22.5% to 57.7% depending only on what wraps it, a lift larger than model switching has typically produced in manipulation benchmarks. The robustness under scene randomization, where learned policies drop by up to 33.9%, reflects the advantage of explicit state tracking over policies fitted to scene-specific distributions. Practitioners comparing model upgrades against harness redesign now have a concrete data point. The zero-reset, zero-demonstration constraint means the approach is testable on physical hardware without task-specific setup.