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. 0:04 – 0:42

    Why robotics is the next frontier for AI beyond software

    The video opens by framing robotics as the clearest bridge from frontier AI’s software strengths into real-world, physical action. The team introduces “Project Fetch” as a practical way to measure how much Claude can accelerate robotics work for non-roboticists.

    • Frontier AI’s impact is well-known in software; the question is how it transfers to the physical world
    • Robotics is positioned as the entry point for software systems to affect reality
    • Project Fetch is designed as a measurable, self-contained experiment
    • Goal: quantify Claude’s acceleration on a sophisticated task for novices
  2. 0:42 – 1:02

    Project Fetch setup: one-day, three-phase benchmark with two teams

    The experiment is defined as a one-day challenge split into three escalating phases, all centered on a robot dog fetching a beach ball. Two comparable teams participate—one with access to Claude and one without—both with minimal prior robotics experience.

    • One-day experiment to keep conditions controlled and comparable
    • Three phases, each framed as “get the robot dog to fetch a beach ball”
    • Two teams: with Claude vs. without Claude
    • Participants are software/research engineers with little robotics background
  3. 1:02 – 1:14

    Phase 1 (manual): using pre-built controllers to perform basic fetch

    Phase 1 tests the simplest baseline: driving the robot dog with provided controllers to retrieve the ball and return it. The teams quickly get hands-on feedback, and the Claude team completes the task faster.

    • Objective: walk to the ball and bring it back using pre-provided controls
    • Early interaction feels intuitive and highlights basic operational learning
    • Claude team completes in ~7 minutes
    • Non-Claude team completes in ~10 minutes
  4. 1:14 – 1:52

    Phase 1 field footage: speed, coordination, and playful competition

    The transcript captures candid team chatter as they navigate where to bring the ball and react to the opposing team’s progress. The segment emphasizes real-time coordination and the messy, human side of physical tasks.

    • Teams negotiate task details (e.g., where to return the ball)
    • Competitive dynamics emerge while both teams operate in the same space
    • Hands-on control reveals small operational frictions quickly
    • Phase 1 serves as a warm-up for later programming-heavy phases
  5. 1:52 – 2:10

    Phase 2 (programmatic): build a custom controller and connect laptop-to-robot

    Phase 2 raises difficulty: teams must write software to control the robot and, critically, establish reliable communication between their laptops and the hardware. The segment highlights how setup and dependency management become the dominant challenge.

    • Task shifts from manual control to writing a controller program
    • Key hurdle: getting a laptop to talk to complex hardware reliably
    • Dependency installation and environment issues slow progress
    • The phase stresses systems integration over simple robot driving
  6. 2:10 – 3:20

    Claude’s advantage in Phase 2: finding libraries, resolving setup, enabling access

    The Claude team benefits from rapid discovery of the right software libraries and installation steps. Claude helps them move from “stuck on setup” to “controlling the dog,” showing the value of AI in navigating undocumented or fragmented tooling.

    • Claude “one-shots” substantial controller scaffolding
    • Helps locate appropriate libraries for the specific robot platform
    • Guides installation and configuration to establish robot access
    • Speeds up the transition from setup to live testing
  7. 3:20 – 3:55

    Phase 2 live testing: fast progress, safety hiccups, and debugging in the real world

    Once control is achieved, the team tests movement and quickly encounters the realities of physical systems—unexpected motion, near-collisions, and the need for emergency stops. The tone shifts from configuration to operational safety and iterative debugging.

    • Robot movement tests reveal speed and control sensitivity
    • Teams actively avoid obstacles (tables, participants)
    • Unexpected behavior prompts calls to turn the robot off
    • Physical debugging differs from purely software iteration
  8. 3:55 – 4:25

    Phase 2 outcome: time-to-complete and the key bottleneck of hardware interfacing

    Results show the Claude team finishing Phase 2 in about 2 hours and 15 minutes, with the largest uplift coming from connection/setup work. The narrator emphasizes that talking to arbitrary hardware is inherently difficult and a major barrier to robotics experimentation.

    • Claude team completes Phase 2 in ~2 hours 15 minutes
    • Biggest uplift: connecting to the robot and establishing reliable communication
    • Hardware interfacing is highlighted as a primary bottleneck
    • Implication: AI help can unlock robotics for non-experts
  9. 4:25 – 5:08

    Non-Claude Phase 2 struggles: dead ends and guided intervention to proceed

    The team without Claude has difficulty finding a working approach, trying multiple paths without success. Organizers ultimately intervene with a known-good strategy so they can continue to later phases, underscoring the cost of tooling uncertainty.

    • Non-Claude team struggles to get camera/robot access working
    • Multiple attempted approaches fail to unlock progress
    • Experiment organizers provide a working baseline strategy
    • Demonstrates how lack of guidance can stall an entire robotics workflow
  10. 5:08 – 5:48

    Phase 3 (autonomous): end-to-end fetch with perception, search, and control

    Phase 3 targets the real long-term challenge: autonomy. The robot must search for the ball, detect it, navigate to it, and return—requiring a full pipeline rather than isolated components.

    • Goal: press ‘go’ and have the robot fetch the ball autonomously
    • Requires perception (detect ball), planning/search, navigation, and control
    • Difficulty is intentionally ratcheted up to reflect real-world autonomy needs
    • Serves as a proxy for what frontier models must solve for useful robotics
  11. 5:48 – 6:17

    Phase 3 progress comparison: partial pipelines vs. near end-to-end integration

    Without Claude, the team makes solid progress on localization and some ball detection but struggles to integrate everything into a working system. With Claude, the team comes close to completion—estimated about 1.5 hours away by the end.

    • Non-Claude team: strong start on tracking robot location in space
    • Non-Claude team: some progress on ball detection, but integration fails
    • Claude team: comes fairly close to finishing the full autonomous task
    • Integration (‘knitting everything together’) is the hardest step
  12. 6:17 – 7:39

    Final results and implications: AI as the bridge to physical-world capability

    The video concludes that Claude enabled teams to complete comparable work hours faster, largely by reducing friction in setup and implementation. It argues that near-term models will let non-experts engage meaningfully with robots, and long-term trends point toward AI systems increasingly doing the full task themselves.

    • Claude team completes their completed tasks a couple hours faster overall
    • Near-term: AI helps novices do meaningful robotics work without specialized training
    • Claude’s benefits emerged without robotics-specific training—an emergent capability
    • Long-term: tasks may shift from human+AI collaboration to primarily AI execution in the physical world

Get more out of YouTube videos.

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