> For the complete documentation index, see [llms.txt](https://docs.roboset.org/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.roboset.org/training-platform/overview.md).

# Overview

### Overview

The training platform captures the complete Physical AI learning loop in the browser:

task attempt → structured episode data → policy training → saved model → evaluation → reward.

You never need hardware. You select a task, generate data through training or demonstration, and the platform turns those attempts into structured episodes, trains a neural-network policy on them, saves the model, evaluates it, and credits your rewards.

### Tasks

Eight manipulation tasks are live, chosen to span the fundamental challenge classes in robotic manipulation. They form a progression from simple transport to precision insertion to multi-step sequencing.

<table data-header-hidden data-search="false"><thead><tr><th></th><th></th><th></th></tr></thead><tbody><tr><td>Task</td><td>Type</td><td>What it teaches</td></tr><tr><td>Pick Stone</td><td>Pick-and-place</td><td>The foundational grasp-and-move primitive</td></tr><tr><td>Move Cube</td><td>Transport</td><td>Moving an object to a marked pad</td></tr><tr><td>Lift Ball</td><td>Lift-and-place</td><td>Adding vertical precision to placement</td></tr><tr><td>Place In Bin</td><td>Transfer</td><td>Placing an object into a container target</td></tr><tr><td>Place Ring</td><td>Precision insertion</td><td>Placing a ring onto a peg</td></tr><tr><td>Drop Token</td><td>Release timing</td><td>Dropping into a target slot</td></tr><tr><td>Stack Books</td><td>Stacking</td><td>Multi-object stability</td></tr><tr><td>Separate Colors</td><td>Sorting</td><td>Sequential multi-placement into matching zones</td></tr></tbody></table>

Each task has a difficulty level, a base reward, visual UI elements, and a backend environment. Additional task names exist in the codebase but are hidden until supported; the feed only shows the current visible set.

### The Robot Environment

Tasks run inside RobotPickEnv, a Gymnasium-compatible environment. This is the same interface standard used by professional reinforcement-learning research, which means everything collected on the platform is architecturally compatible with what labs actually train on.

**Observation** — 20 dimensions. The environment returns a 20-value observation vector at every timestep describing the full robot and task state: robot joint information, end-effector / gripper position, object position, target position, gripper open/closed state, and distance measurements to object and target.

**Action** — 8 dimensions. The policy outputs an 8-value continuous action vector: seven joint controls plus gripper control. The environment applies the action, updates the state, computes a reward, and returns the next observation.

**Physics**. The environment can use PyBullet physics when available and falls back to a deterministic simulator otherwise. The fallback keeps the platform usable on any host regardless of configuration. By default the environment favors the fallback unless PyBullet is explicitly enabled and imports successfully.

**Reward shaping**. Actions are scored so that moving toward the object, achieving a grasp, placing correctly, and completing the task yield progressively better reward. This shaping is what lets a policy learn from sparse task success.

<br>

<br>


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.roboset.org/training-platform/overview.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.