Training Robotics Models - Part 1: Buy vs Build

A New Series - Trelis Robotics

A new Trelis series on robotics!

In this video, I walk through the SO-101 robot and talk about the decision to buy vs build.

I also give a quick preview of the main robotics models, ACT and GR00T-N1 by Nvidia, as well as pi0.

Cheers, Ronan

Timestamps:

0:00 Building vs Buying an SO-101 Robot

0:38 Video Series Overview (Build/Buy, Models, Training, Fine-tuning)

1:33 Today’s Video Overview: SO-101; Build v Buy; Robot Models

2:26 Leader-Follower Setup

4:35 SO-101 Overview

5:47 SO-101 vs SO-100 vs Moss v1

6:43 How a Robot AI Model works

11:48 How are Robot AI Models different from LLMs?

13:50 The main types of robot models (ACT, GR00T, pi0)

16:05 Building vs Buying a Robot

19:01 3D Printing and Printer Tips

20:09 Purchase the Github Repo: Trelis.com/ADVANCED-robotics

---

Trelis Links:

🤝 Are you a talented developer? Work for Trelis

💡 Need Technical or Market Assistance? Book a Consult Here

💸 Starting a New Project/Venture? Apply for a Trelis Grant

---

Building vs. Buying Your First Robot for AI Training

A practical guide to getting started with robotic learning systems, focusing on the SO-101 robot platform and key technical considerations.

Robot Overview: The SO-101 System

The SO-101 consists of two robots:

1. A leader robot for demonstration and data collection

2. A follower robot that learns to replicate tasks

Key components:

1. 6 motors controlling different joints

2. Wrist camera mounted on the robot arm

3. Front-view camera capturing the full workspace

4. PID control system for motor position matching

The SO-101 model improves on the SO-100 with:

1. Better motor wiring organization

2. Specialized motors with looser gearing for easier manual movement

3. No need to remove gears for leader robot operation

Technical Architecture

Core system elements:

1. Dual camera inputs (wrist + front view)

2. 6 motor position sensors

3. Neural network predicting future motor positions

4. 10-step prediction horizon (~200ms) for each motor

5. PID control loop executing predicted positions

Key differences from language models:

1. Non-autoregressive prediction

2. Scalar outputs (motor angles) vs discrete tokens

3. Minimal historical context

4. Real-time performance requirements

5. Smaller model sizes for speed

Build vs Buy Decision Factors

Build Option:

1. Individual motor costs from Alibaba

2. Assembly and calibration time

3. Total parts cost similar to pre-built

Buy Option:

1. Complete SO-101 system ~200 EUR plus shipping

2. Pre-calibrated motors

3. Faster setup time

4. Includes all core components

Recommended 3D printers if building:

1. Bamboo A1 (256x256mm) - can print all parts at once

2. Ender V3 SE (220x220mm) - requires multiple print batches