Skip to content
AnthropicAnthropic

Who let the robot dogs out?

We asked two teams of Anthropic researchers to program a robot dog. Neither team had any robotics expertise—but we let only one team use Claude. In the past, we’ve run simulated studies where Claude trained a robot dog. These helped us assess how Claude might contribute to AI research and development. Project Fetch was us trying something similar in practice. This project suggests that we’re not far from a world where frontier AI models can interact with previously-unknown pieces of hardware, even with non-experts at the helm. For more, read on: https://www.anthropic.com/research/project-fetch-robot-dog 0:00 Introduction: Why robotics? 0:30 The experiment 1:02 Phase 1: Fetch, manually 1:51 Phase 2: Fetch, programmatically 5:08 Phase 3: Fetch, autonomously 6:24 Results

Nov 12, 20257mWatch on YouTube ↗

CHAPTERS

  1. Why robotics is the next frontier for AI impact

    The video frames robotics as the clearest path for frontier AI models to extend their usefulness beyond software into the physical world. It sets up the motivation: understanding how much an AI assistant can accelerate real-world technical work, not just coding tasks.

    • Frontier AI’s impact has focused on software engineering so far
    • Robotics connects software systems to physical action in the real world
    • Goal: measure how AI assistance translates into physical-world task performance
    • Premise: AI could reduce the expertise barrier to working with robots
  2. Project Fetch setup: a one-day, three-phase experiment

    Project Fetch is introduced as a self-contained, time-boxed test: get a robot dog to fetch a beach ball with increasing levels of difficulty. The structure is explicitly comparative, designed to quantify speed and progress differences with and without Claude.

    • One-day experiment designed for measurable outcomes
    • Three phases, all centered on “robot dog fetches a beach ball”
    • Difficulty increases from manual control to full autonomy
    • Focus is acceleration of unfamiliar but sophisticated technical tasks
  3. Teams and fairness: same background, different tooling

    Two teams of Anthropic software/research engineers—each with little robotics experience—compete under similar conditions. The main variable is whether the team has access to Claude as an assistant.

    • Two teams: one with Claude access, one without
    • Participants are engineers with minimal robotics experience
    • Comparison aims to isolate the effect of AI assistance
    • Measures include time-to-complete and degree of completion
  4. Phase 1 — Manual fetch using provided controllers

    The first task is straightforward: use existing, pre-provided controllers to drive the dog to the ball and return. The segment shows quick onboarding, playful competition, and the first time comparison between teams.

    • Objective: walk to the beach ball and bring it back manually
    • Uses pre-built controllers (no programming required)
    • Team with Claude completes faster (~7 minutes)
    • Team without Claude finishes in ~10 minutes
  5. Phase 2 — Building a custom controller and connecting to hardware

    The challenge shifts from driving the robot to creating a programmatic controller and establishing reliable laptop-to-robot communication. The narrative emphasizes that integration and setup—not just coding logic—become the dominant difficulty.

    • Objective: program a controller from a laptop to operate the robot
    • Requires accessing hardware interfaces and using SDK/tooling (e.g., ROS 2)
    • Dependency hell and missing packages become immediate blockers
    • Core bottleneck: getting a complex robot to “talk” to your computer
  6. Claude’s biggest uplift: troubleshooting setup, libraries, and access

    Claude helps the assisted team rapidly find appropriate libraries, install dependencies, and reach basic control and sensor access. The video highlights how AI collapses the time spent on search, configuration, and “nitty-gritty” details that stall novices.

    • Claude finds relevant software libraries for the specific robot
    • Guides installation of correct packages and dependencies
    • Speeds up gaining access to robot control and camera feeds
    • Illustrates reliance on AI for tedious troubleshooting work
  7. When robots go wrong: safety hiccups and messy real-world control

    As teams gain control, the robot behaves unpredictably at times, leading to near-collisions and comedic panic. This underscores that physical systems introduce risk, latency, and failure modes that don’t appear in pure software demos.

    • Robot control can be unstable or surprising during early setup
    • Real-world constraints: obstacles, collisions, and emergency stops
    • Physical testing introduces safety considerations and interruptions
    • Highlights the gap between “code works” and “robot behaves safely”
  8. Phase 2 outcomes: assisted team finishes; unassisted team stalls and needs intervention

    The Claude team completes Phase 2 in about 2 hours and 15 minutes, largely due to faster hardware connectivity and tooling setup. The team without Claude struggles to find a working approach and ultimately needs a recommended strategy to proceed.

    • Team with Claude completes Phase 2 in ~2h15m
    • Largest time savings: connecting to and communicating with the robot
    • Unassisted team explores multiple unproductive paths
    • Organizers intervene with a known-working strategy to unblock them
  9. Phase 3 — Full autonomy: press go, search, detect, fetch

    The final phase requires an end-to-end autonomous system: the robot must locate the ball, navigate to it, and return without manual input. This phase is framed as closer to the real future challenge—robots executing tasks on behalf of models autonomously.

    • Objective: autonomous search, ball detection, approach, and retrieval
    • Requires integrating perception, localization, planning, and control
    • Represents the “real problem” frontier models may need to solve
    • Difficulty intentionally ratchets toward end-to-end autonomy
  10. Phase 3 progress: localization and detection vs. full integration

    The unassisted team makes meaningful progress on parts like tracking the robot’s position and working on ball detection, but can’t combine components into a complete system. The Claude team gets much closer, nearing completion but not fully finishing within the day.

    • Unassisted team: progress on localization and some ball detection
    • Main failure point: integrating subsystems into a working pipeline
    • Assisted team: comes fairly close; estimated ~1.5 hours from done
    • Demonstrates AI advantage in stitching together complex systems
  11. Results and takeaway: AI speeds robotics work even without robotics-specific training

    Overall, the Claude team completes what they complete a couple of hours faster than the team without Claude. The video stresses that this uplift wasn’t from training Claude specifically for robotics—it emerged from general assistant capabilities applied to a new domain.

    • With-Claude team is “a couple of hours faster” overall on completed work
    • Most decisive advantage appears in setup/integration steps
    • Claude wasn’t specialized for this experiment; benefit is emergent
    • Suggests AI can enable non-experts to work meaningfully with robots
  12. Forward-looking implication: AI’s effects will extend from software into hardware and autonomy

    The conclusion argues that near-term AI will help humans interface with robots more easily, and longer-term, tasks may shift from human+AI collaboration to AI-only operation. The broader claim is that frontier AI will increasingly influence the physical world, not just software.

    • Near-term: AI lowers the barrier for people to use robotics systems
    • Long-term: what needs a person+model today may need only the model tomorrow
    • Robotics is a leading indicator of AI moving into physical domains
    • AI impact expected to span both hardware and real-world operations

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.