Skip to content
Get packs

Guide · Agent skill

How to use the Robotics Idea Discovery agent skill (idea-discovery-robot)

Updated · written with AI from the project's own documentation and checked against our catalogue data

The Robotics Idea Discovery skill (idea-discovery-robot) is a workflow for Claude-style coding agents that takes a broad robotics direction, such as manipulation or drones, and turns it into a short list of testable research ideas. It is written for robotics and embodied AI researchers, students planning a robot learning project, and teams that want a simulation-first pilot plan. This guide explains what the skill does, how to install it, how to trigger it, what it needs, and where to be careful.

Robotics Idea Discoverywanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Install the skill. Download the ZIP from the page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
  2. 2Install the sub-skills. Make sure /research-lit, /idea-creator, /novelty-check and /research-review are available, since the pipeline calls them.
  3. 3Set up the external reviewer. Configure Codex MCP (mcp__codex) with the gpt-6-astra model, or configure another reviewer.
  4. 4Frame your robotics direction. Choose a direction such as manipulation, locomotion, navigation, drones or humanoids, and add any constraints you have.
  5. 5Trigger the skill. Ask in natural language, such as "robotics idea discovery for manipulation", or run /idea-discovery-robot with your direction.
  6. 6Review the checkpoints. Read the landscape summary and the shortlisted ideas, and tell the agent to change scope if needed.

What does the idea-discovery-robot skill make the agent do?

An agent skill is a folder with a SKILL.md file that tells an AI agent how to carry out a specific workflow. This skill makes the agent run a multi-step pipeline: literature survey, idea generation, novelty check, and critical review.

The skill frames every phase around robotics details such as embodiment, benchmarks, observation and action interfaces, and sim-to-real constraints. It favours simulation and offline data over real hardware.

  • Phase 0: frame the problem (embodiment, task family, environment, modalities, action interface, assets, safety constraints)
  • Phase 1: robotics-focused literature survey, grouped by embodiment, benchmark, action interface and sim2real setup
  • Phase 2: idea generation and filtering with failure metrics and baseline clarity required
  • Later phases: novelty check and critical review of the shortlisted ideas

How do you install the idea-discovery-robot skill?

Download the skill ZIP from the page, then unzip it so the idea-discovery-robot folder (containing SKILL.md) sits in one of the skills locations below. For a personal install, use ~/.claude/skills/. For a single project, use the project's .claude/skills/ folder.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

  • Personal install: unzip into ~/.claude/skills/
  • Project install: unzip into the project's .claude/skills/
  • Claude apps: Settings → Capabilities → Skills, then upload

How do you trigger the skill with a natural request?

The skill activates when your request matches its description. You can say things like "robotics idea discovery", "embodied AI idea", "sim2real 选题", or ask for ideas on manipulation, locomotion, navigation, drones, humanoids, or general robot learning.

You can also call it directly with a robotics direction as the argument. Inline overrides are supported, for example limiting the run to simulation-only ideas or to a particular venue focus.

  • "Find me robotics research ideas for bimanual manipulation"
  • "Embodied AI idea for drone navigation, focus on CoRL or RSS"
  • "sim2real 选题 for legged locomotion"
/idea-discovery-robot "bimanual manipulation" — only sim ideas, no real robot

What examples show the skill in use?

The source gives these override patterns. Each one changes the defaults for a single run.

The skill's default constants include at most 3 ideas validated deeply, simulation-first pilots, and no real-robot pilots without explicit approval.

  • Simulation-only run: "bimanual manipulation" with only sim ideas and no real robot
  • Venue-focused run: "drone navigation" aimed at CoRL/RSS with 2 pilot ideas maximum
/idea-discovery-robot "drone navigation" — focus on CoRL/RSS, 2 pilot ideas max

What does the skill need to run?

The skill depends on other skills and tools. It will not produce a full pipeline on its own, because the sub-skills it calls are not defined in this file.

  • Claude Code or a compatible agent, with the sub-skills /research-lit, /idea-creator, /novelty-check and /research-review installed
  • Codex MCP (mcp__codex) for the default external reviewer, with model set to gpt-6-astra, unless another reviewer is configured
  • Web search and fetch access for literature retrieval

What are the safety notes for scripts and network access?

This skill runs scripts and uses the network. Its allowed tools include Bash, file read and write, search, web search and fetch, and Codex MCP calls. Review what the agent runs and fetches, especially if your environment holds private data.

The skill is designed not to execute on real hardware unless you explicitly approve it. Keep real-robot pilots out of the run unless you have decided otherwise.

  • Runs scripts: yes
  • Uses network: yes (literature retrieval and external review)
  • Real-robot pilots require explicit approval; the default is sim-first or offline-log pilots

What tips help you get useful output?

Give the skill a specific robotics direction rather than a broad field. Name the embodiment, benchmark family or constraints where you know them, because the frame drives every later phase.

Expect checkpoints. The skill proceeds automatically if you do not respond, using the strongest sim-first option, so tell it to stop or narrow scope if you want to steer.

  • Name the embodiment, task family and benchmark you care about
  • State constraints up front, such as no hardware or a specific venue
  • Treat the novelty result as a lead to check, not a verdict

What people use it for

  • Choosing a thesis topic in robot learning with a benchmark-grounded starting point
  • Scoping a simulation-first pilot before booking hardware time
  • Surveying recent CoRL, RSS, ICRA, IROS and RA-L work on a manipulation or locomotion topic
  • Identifying recurring failure modes that papers in a task family do not address

Frequently asked questions

Is the idea-discovery-robot skill free to use?

The source file is published under the MIT licence, which the item listing states. Check the licence file in the repository for the full terms. The skill also depends on other tools and services, such as Codex MCP, that have their own access requirements.

What do I need to install idea-discovery-robot?

You need Claude Code or a compatible agent, the sub-skills /research-lit, /idea-creator, /novelty-check and /research-review, and web search and fetch access. For the default reviewer, you also need Codex MCP with the gpt-6-astra model unless you configure another reviewer.

Where do I put the idea-discovery-robot skill folder?

Unzip the downloaded ZIP into ~/.claude/skills/ for a personal install, or into a project's .claude/skills/ folder for a single project. In the Claude apps, upload it under Settings → Capabilities → Skills.

Does the skill run real robot experiments?

No, the skill is designed to avoid hardware execution unless you explicitly approve it. Its default pilots are simulation-first or based on offline logs.

Can the novelty check be trusted?

Not as a verified judgement. The novelty check relies on search and model review, which can miss prior work, so verify any idea against the literature yourself before committing time to it.

How is this different from a general research idea tool?

This skill frames ideas around robot details such as embodiment, benchmarks, action interfaces and sim-to-real constraints. A generic ML idea tool does not apply those robotics-specific checks by default.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you are a robotics or embodied AI researcher, a student planning a robot learning project, or a team that wants a simulation-first pilot plan before using hardware. Pick something else if you do not have the sub-skills and reviewer setup it depends on, or if you need a verified novelty judgement.

Download the skill