AI for Robotics Control and Embodiment

4 posts · Nov 2025 – Jul 2026

What this trail covers

Anthropic is exploring and evaluating the capabilities of large language models in controlling robotic systems. This involves testing models across various robotic platforms (simulated and real) and control interfaces, ranging from low-level motor torque commands to high-level policy supervision. The research aims to understand how LLMs can synthesize logical skills and 3D understanding for physical world interaction, and to identify the most effective control mechanisms for achieving reliable robot operation. Recent experiments with robot dogs demonstrate significant AI uplift in task completion speed and success rates, bridging the digital and physical worlds.

All Anthropic trails

Jul 2026Project Pilot: Can AI models fly drones?This post introduces Project Pilot and the Drone-Bench benchmark, evaluating AI models' ability to autonomously control a drone for a locate-and-follow surveillance task. It details the decomposition of the task into five sub-tasks: Reconstruct, Localize, Navigate, Detect, and Follow. The post presents evaluation results for various models, highlighting Claude Fable 5's performance and its specific failure in the reconstruction sub-task. It also discusses the importance of consistency in model performance and the implications of AI's increasing capability in hardware operation for both opportunities and risks, drawing parallels to agentic coding.
Jul 2026How Claude Performs on Robotics TasksThis post details Anthropic's evaluation of language models' performance in robotics across various control interfaces and robot bodies. It highlights that model performance is highly dependent on the control interface, with models struggling with direct low-level torque control but showing significant gains when supervising pretrained policies or using high-level commands. The research quantifies generational improvements in locomotion and manipulation tasks, noting that newer models are better at adapting strategies and converting sensory input into actions. The post also identifies limitations, such as difficulties with stable spatial memory and long open-loop plans, and discusses the implications for safe AI development.
Jun 2026Project Fetch: Phase twoThis post details the second phase of Project Fetch, demonstrating significant advancements in AI's ability to autonomously control robotic systems. Claude Opus 4.7 achieved task completion speeds up to 20 times faster than previous human teams, and over 18 times faster than human teams assisted by Claude Opus 4.1. The model generated more effective code with less volume, and demonstrated improved identification of optimal interfacing paths. While the model still struggles with precise manipulation tasks like beach ball retrieval, it shows a clear trajectory towards autonomous physical agentic AI, mirroring previous advancements in AI-assisted software engineering.
Nov 2025Project Fetch: Can Claude train a robot dog?This post details an experiment using Claude to assist researchers in programming a robot dog to fetch beach balls. The experiment demonstrated a significant 'uplift' in task completion speed and success rate for the team with Claude access compared to the team without. Key technical contributions include: quantifying AI uplift in a physical robotics task, highlighting challenges in hardware connection and sensor data utilization, and observing how AI assistance can lead to broader exploration and potentially more code, but also more efficient problem-solving. The experiment also provided qualitative insights into team morale and dynamics when using AI assistants.
Advertisement