---
title: Auto-Curate Your Data
sidebar:
  order: 28
---
Robotics teams collect far more data than anyone can review by hand. The interesting moments — a hard landing, a stair climb, a near-miss — are buried in hours of normal operation. This guide shows how to go from raw logs to a curated set of events automatically: you describe *what qualifies* as an interesting moment once, and Roboto's AI finds every occurrence and organizes them into a [Collection](/docs/learn/collections) for you.

By the end of this guide you will have:

1. A [Skill](/docs/learn/ai/skills) that captures how to detect a **Stair Climb Event** from the elevation signal and the front camera topic.
2. An [Agent](/docs/learn/ai/agents) that applies the skill to any dataset and — through a [Goal](/docs/learn/ai/goals) — is *required* to create the events it finds and add them to a collection.
3. A Collection of validated Stair Climb Events, ready for downstream review or conversion into an ML-ready format like [LeRobot](https://github.com/huggingface/lerobot).

We'll use two example logs from the [GrandTour Dataset](https://grand-tour.leggedrobotics.com/), recorded on an ANYmal legged robot.

## Preparation

- Create a [Roboto Account](/docs/getting-started/account) and sign in.
- Learn the core concepts of Roboto: [Concepts](/docs/learn/concepts).
- This guide assumes your organization has a ROS ingestion trigger configured, so uploaded `.mcap` files are ingested automatically. If yours does not, follow [ROS Triggers](/docs/user-guides/working-with-ros-logs/trigger-ros) first.
- Download the two sample logs: [arche\_small.mcap](https://roboto-ros-demo-datasets.s3.us-west-2.amazonaws.com/arche_small.mcap) and [polyterrasse\_small.mcap](https://roboto-ros-demo-datasets.s3.us-west-2.amazonaws.com/polyterrasse_small.mcap).
- Using the **\+** icon in the top navigation bar, create a dataset and upload `arche_small.mcap` to it. Then create a second dataset for `polyterrasse_small.mcap`. We'll use the first dataset to develop and test the skill, and the second to run the agent unattended.
- Roboto ingests the `.mcap` files automatically on upload. Once ingestion completes, each dataset has the two topics this guide uses: the state estimator (`/anymal/state_estimator/anymal_state`) and the front camera (`/boxi/hdr/front/image_raw/compressed`).

## Create a Skill

A [Skill](/docs/learn/ai/skills) is a stored, versioned procedure for Roboto's AI. Writing the stair-climb detection logic as a skill means everyone on your team — and every agent — can run the same investigation.

1. Go to **Agents → Skills** in the web app and click **Create Skill**. Name it `detect_stair_climb` and fill in the two fields:

   **When to use** — the short description the AI reads when deciding whether the skill applies to a request:

   ```text
   Apply this skill to determine if a stair climb can be observed in provided data.
   ```

   **Procedure / body** — the instructions the AI follows:

   ```text
   Find instances of stair climbing.

   First, look at [[/anymal/state_estimator/anymal_state.contacts.position.z]]
   and find periods where there is a significant elevation change (>0.4m).

   Second, look at the [[/boxi/hdr/front/image_raw/compressed]] images and
   validate that the robot is climbing stairs.

   For each validated stair climb, create a Roboto Event.
   ```

   Type `[[` in the editor to reference a topic — referenced topics are pre-loaded for the AI, so it starts the investigation already knowing where to look.

   ![The Skill editor showing the detect\_stair\_climb skill, with a "when to use" description and a procedure that references the state estimator and front camera topics](/docs/blume-assets/content/docs/_static/agents-skill-editor.png)

Note how the procedure combines two signals: a numeric threshold on the elevation channel to find *candidates*, and the camera to *validate* that a candidate is a staircase, not a ramp or slope. Being explicit about the signals and the decision rule makes a skill reliable; vague instructions leave the AI room to guess.

:::tip
You can also create a skill interactively: explore a dataset in AI Chat until the investigation works, then ask the AI to save the conversation as a skill. It drafts the name, description, and body for your review — nothing is saved until you approve it.
:::

## Test the Skill on the First Dataset

2. Open `arche_small.mcap` from the first dataset in the visualizer, and open **AI Chat** from the sidebar on the right. Type `/` in the chat composer and select `detect_stair_climb` to invoke the skill on this file.

   ![The visualizer showing the front camera and elevation plot for arche\_small.mcap, with the AI Chat panel open and the skill picker showing detect\_stair\_climb after typing "/"](/docs/blume-assets/content/docs/_static/agents-invoke-skill.png)

3. The AI works through the procedure: it analyzes the elevation channel for candidate intervals, samples camera frames to confirm the robot is on a staircase, and creates a **Stair Climb Event** for the confirmed interval. When it finishes, the event appears on the timeline, and the chat explains how it was detected — the threshold crossing, the visual confirmation, and the event boundaries.

   ![The visualizer after the skill has run, showing a Stair Climb Event highlighted on the elevation plot, the event details panel, and the AI's explanation of how the event was detected](/docs/blume-assets/content/docs/_static/agents-skill-result.png)

The event is now part of the dataset — you can inspect its time bounds against the camera and plot, adjust it, or delete it like any other [event](/docs/learn/concepts#events-section).

## Create an Agent

Testing the skill in chat works well for a single dataset, but doesn't scale to every new recording. That's what an [Agent](/docs/learn/ai/agents) is for: a named, reusable AI worker that binds instructions, skills, and a required [Goal](/docs/learn/ai/goals) to typed variables — so anyone on your team can run the same curation job on a new dataset with one click.

4. Go to **Agents → Create Agent** and fill in the **Details** section:

   - **Name**: `Stair Climb Agent`
   - **Description**: `A demo agent that curates Stair Climb Events based on the robot's position and front camera.`
   - **Model**: `Standard`
   - **Kickoff message**: `You are tasked to analyze legged robot data and identify Stair Climb Events.`

   ![The Details section of the agent form, with the name, description, model, and kickoff message filled in](/docs/blume-assets/content/docs/_static/agents-agent-form-details.png)

5. In the **Goals** section, add a **Create events** goal. A goal is a *required* outcome: the agent's turn succeeds only after the goal is achieved, making the workflow dependable enough to run unattended.

   - **Dataset**: bound to `{{dataset.id}}` — filled in at launch time.
   - **Event types**: one entry, `Stair Climb Event` with the description `When a robot ascends stairs.` The agent can only create events of the types you declare here.
   - **Add created events to a collection**: enabled. The collection is bound to `{{collection.id}}`, also filled in at launch time.

   ![The Goals section of the agent form, showing a Create events goal with a Stair Climb Event type, bound to dataset and collection variables](/docs/blume-assets/content/docs/_static/agents-agent-form-goal.png)

6. In the **Skills** section, attach `detect_stair_climb` so the agent investigates datasets using the exact procedure you tested. Pin a version, or leave it on **Latest** to track the newest.

   ![The Skills section of the agent form, with detect\_stair\_climb attached and its version set to Latest (v1)](/docs/blume-assets/content/docs/_static/agents-agent-form-skills.png)

Save the agent.

:::tip
Like skills, agents can also be created from a conversation: ask AI Chat to turn the current thread into a reusable agent, then review and save the pre-filled draft.
:::

## Create a Collection for the Events

7. Go to **Collections**, click **\+ New**, and create an empty collection to receive the curated events:

   - **Name**: `stair_climb_collection`
   - **Description**: `Collection with auto-curated stair climb events`
   - **Entity type**: `Event`

   ![The New collection dialog on the Collections page, creating an empty Event collection named stair\_climb\_collection](/docs/blume-assets/content/docs/_static/agents-create-collection.png)

Copy the new collection's ID (it starts with `cl_`) — you'll need it at launch.

## Launch the Agent on the Second Dataset

8. Open the **Stair Climb Agent** and click **Launch**. The launch form asks for the agent's two variables: the collection you just created, and the second dataset (the one holding `polyterrasse_small.mcap`) — a dataset the agent has never seen.

   ![The agent launch form with the collection.id and dataset.id variables filled in](/docs/blume-assets/content/docs/_static/agents-launch-form.png)

9. Click **Launch**. This starts a [thread](/docs/learn/ai/agents) you can follow live: the task list on the left tracks the agent's progress as it loads the skill, analyzes the elevation data, validates candidates against the camera, and creates the events. On this log it finds and creates two Stair Climb Events, each with its methodology spelled out — and because of the goal, both are added to your collection before the turn can complete.

   ![The agent thread view, showing the completed task list and the agent's summary of two created Stair Climb Events with their time bounds, elevation gains, and detection methodology](/docs/blume-assets/content/docs/_static/agents-thread-progress.png)

## Review the Results

10. When the thread finishes, open `stair_climb_collection`. It now contains the two events, each carrying a description of what happened and why it qualified.

    ![The stair\_climb\_collection page listing two Stair Climb Events, with one expanded to show its description, time bounds, and dataset scope](/docs/blume-assets/content/docs/_static/agents-collection-events.png)

11. Click through an event to validate it in the visualizer: the event interval is highlighted on the elevation plot, and the camera confirms the robot on the staircase at that moment.

    ![The visualizer on polyterrasse\_small.mcap showing a created Stair Climb Event highlighted on the elevation plot, with the front camera at the staircase and the event details panel open](/docs/blume-assets/content/docs/_static/agents-validate-event.png)

## Creating Agents from the SDK

Everything above — creating skills and agents, and launching them with goals — can also be done programmatically. See [Agents and Threads](/docs/learn/ai/agents), [Skills](/docs/learn/ai/skills), and [Goals](/docs/learn/ai/goals) for SDK examples.

## Conclusion

You went from two raw logs to a curated, versioned collection of validated events — and the judgment that produced it lives in a skill and an agent your whole team can reuse on the next dataset.

From here:

- **Curate at scale**: launch the agent on any new dataset; every confirmed event lands in the same collection.
- **Hand off downstream**: collections are versioned, so you can pin a version and hand the exact event set to a reviewer or a training pipeline. For example, you can use an action from the [Action Hub](https://app.roboto.ai/hub) to convert the curated events into a [LeRobot](https://github.com/huggingface/lerobot) dataset on Hugging Face.
- **Automate fully**: pair the agent with an [Action](/docs/learn/actions) and a trigger to launch it automatically whenever a new dataset finishes ingestion.