Claude Opus 4.7 controls robodogs 20x faster
Anthropic's Frontier Red Team published Project Fetch Phase Two findings, showing that Claude Opus 4.7 autonomously completed robodog sensor and coding tasks 20 times faster than human teams. While the model wrote ten times less code to interface with hardware, it still struggled with closed-loop physical control like steering a beach ball.
While autonomous code generation and sensor integration speedups are impressive, the physical "last mile" of robotics control remains a hard barrier for pure LLMs lacking real-time closed-loop control policies.
* Rapid Translation of Intent to Code: Claude Opus 4.7 completed in under 10 minutes (averaging 12 minutes across all five tasks) what took human teams hours, proving that LLMs can near-instantly interface with unfamiliar hardware APIs.
* Drastic Code Volume Reduction: The autonomous model generated roughly 1,045 lines of code—nearly a tenth of the code written by the human-AI pair-programming team (10,309 lines)—indicating high programmatic efficiency.
* The Closed-Loop Bottleneck: Despite superior programming efficiency, the lack of real-time sensorimotor adaptation prevented Claude from completing the physical "fetching" task, showcasing the divide between code generation and physical actuation.
DISCOVERED
48d ago
2026-06-18
PUBLISHED
48d ago
2026-06-18
RELEVANCE
AUTHOR
AnthropicAI