GPT-6-Astra tests reveal persistent sim-to-real gap in fine motor robotics

New benchmark evaluations for GPT-6-Astra show that high simulation scores fail to translate into reliable fine motor robotics performance.

The evaluations assessed the model across twelve robotics benchmark suites designed to test autonomous physical control. GPT-6-Astra achieved strong marks within synthetic training environments, demonstrating stable trajectory planning and multi-step reasoning. However, performance degraded significantly when the same policies were deployed on physical hardware.

The divergence proved most pronounced during fine motor manipulation tasks. In simulated environments, virtual agents executed precise insertion, tactile grasping, and thin-object manipulation with consistent success. On physical test benches, identical command sequences failed to account for real-world physical dynamics.

Hardware trials exposed recurring vulnerabilities to unmodelled friction, surface compliance, and mechanical micro-slippage. Tasks requiring fine force modulation, such as threading components or handling deformable materials, suffered the sharpest declines. While the model could plan macro-level arm movements effectively, it struggled to maintain the millisecond-level tactile adjustments needed to stabilise delicate interactions.

The results highlight the ongoing limitations of purely simulated training distributions. In synthetic environments, policy rollouts benefit from deterministic contact mechanics and noise-free state measurements. Real hardware introduces sensor latency, calibration drift, and tactile contact noise. GPT-6-Astra frequently compounded these minor sensor errors, leading to missed grasp points or excessive contact force.

The findings indicate that scale alone has not eliminated the simulation-to-reality boundary for dexterous manipulation. While massive multimodal training expands high-level problem-solving capabilities, physical execution still demands grounded, high-frequency closed-loop feedback. As robotics labs attempt to move large foundation models into production manufacturing and laboratory automation, bridging this physical discrepancy remains the primary technical bottleneck.

If synthetic environments cannot replicate sub-millimetre physical interactions, how much useful guidance can purely simulated scaling continue to offer real-world robotics?