A new framework called HarnessPAI wraps robot control models in evolvable code, and early results suggest it fixes some of the blind spots that plague today's action-only robot AI.
The work comes from a paper posted to arXiv on September 25, 2026 (arXiv:2609.29166, https://arxiv.org/abs/2609.29166). The researchers built a harness that separates robot behavior into two timescales: within a single task attempt, it runs a fixed program at the code level, open-loop, to guide and check execution; across many attempts, it revises that program using execution feedback and turns failures into reusable skills. Tested across desktop robot arms, household robots, a robot vacuum, and a legged walking agent, the harness beat both raw action models and prior code-as-policy setups without retraining the underlying model - a 61.6-point gain over the pi0.5 action model on the LIBERO-PRO benchmark and a 27.2-point gain over the WorldDreamer baseline on RoboCasa's atomic tasks. The team also used a converged program to collect expert data, and fine-tuning pi0.5 on that data lifted its own LIBERO-PRO success rate by another 38.8 points.
That's a pointed critique of where physical AI research has been spending its effort. Most recent progress, including action models like pi0.5, has focused on turning what a robot sees into motor commands, while perception and reasoning get treated as solved. This paper argues that's backwards: without a system that also checks and corrects what the robot understands about a scene, even a strong action model breaks down on long, multi-step tasks or scenes that stray from its training data.
It's a variation on the code-as-interface idea that's circulated in robotics labs for years, but letting the program itself evolve from failures is what produced the bigger numbers here. Worth remembering: these are benchmark gains on LIBERO-PRO and RoboCasa, not a robot working unsupervised in your kitchen.