# The Physical AI Sprint — full handbook

Generated from the event handbook at https://www.buildspace.tv/. Event date 2026-08-17.

## The event

A one-day hackathon at the intersection of AI and the physical world, alongside Actuate SF.
Hosted by Nebius with NVIDIA, Antioch, and Toloka. 100-125 participants in teams of 1-5 (solo is fine).
Free to attend; applications are approved because space is limited.

### The challenge

Modern AI excels at generating text and pixels. Physical AI requires closing the loop
between the digital mind and a physical embodiment, integrating three pillars:

- **Perception (sense)** — interpret camera and sensor data into a live understanding of a
  scene: RGB-D video, LiDAR, IMU readings, tactile feedback, spatial audio.
- **Reasoning (think)** — turn goals into multi-step behavior and adapt when the world
  changes, using foundation models, spatial computing, and decision-making frameworks.
- **Action (act)** — translate plans into robot motion that actually works: manipulation,
  autonomous navigation, drone flight dynamics, or IoT actuation.

Three project directions, all equally viable: **sim only** on the Antioch platform,
**hardware only** on the physical robots, or **sim and real** — prototype in sim and run on
the real robot, or capture real demonstrations and bring them into sim. Hardware is limited
and shared across all teams, so simulation is the dependable path.

### Schedule (subject to change)

| Time | What |
| --- | --- |
| 8:00 am | Doors open, check-in, breakfast |
| 9:00 am | Welcome and challenge briefing |
| 9:15 am | Technical workshops by hosts |
| 10:30 am | Hacking begins |
| 12:00 pm | Lunch |
| 1:00 pm | Hacking continues |
| 3:30 pm | Submission deadline |
| 4:30 pm | Demos and judging (top 6 live demos) |
| 5:30 pm | Winners announced |
| 6-9 pm | Happy hour |

### What you get

- **The Antioch platform** — cloud simulation for robotics. Write ordinary Python against
  the NVIDIA Isaac stack (Isaac Sim and Isaac Lab both available) on your own laptop;
  Antioch dispatches it to GPU machines on Nebius. No local GPU, Docker, or Isaac install.
- **Robot hardware** — 10 LeRobot SO-101 leader/follower arm pairs, 2 Unitree Go2-W wheeled
  quadrupeds, 1 Unitree G1 humanoid. Stations are shared and staffed by the host teams.
- **Sim-ready assets** — a calibrated SO-101 digital twin, the Unitree lineup in the Isaac
  asset library, ready-made environments, and props.
- **Physical assets** — blocks of various size, shape, and color; vials and vial racks.
- **Frontier models** — NVIDIA Nemotron and Cosmos Reason over an OpenAI-compatible API.
  Start on Nebius Token Factory (https://api.tokenfactory.nebius.com/v1/), where new
  accounts get $25 in credits plus $25 in Tavily credits; NVIDIA's own endpoint
  (https://integrate.api.nvidia.com/v1) serves what Token Factory does not, Cosmos Reason
  above all. Cosmos Reason is a vision-language model that reasons about the physical
  world from video and images; Nemotron covers planning and tool calling.
- **Support** — engineers from the host companies on hand all day.

### Judging

Science-fair style first: judges circulate and view projects at team stations, so have a
tight walkthrough and a live or recorded demo ready. The top 6 teams then demo to the full
group. Four unweighted categories:

- **Ambition** — how hard the problem is and how much of the loop you take on.
- **Functionality** — does it work end-to-end when demonstrated.
- **Creativity** — originality in the task, approach, or demo.
- **Architectural quality** — clean boundaries between perception, reasoning, and action.

### Ground rules

Unless you are confident you can pull it off quickly, avoid fine-tuning pre-trained
policies or VLA models. Training a small imitation policy from scratch on your own
demonstrations fits comfortably in the day; wrangling a large pre-trained model usually
does not. Make it awesome.

### Who should attend

Robotics engineers, machine learning engineers, computer vision engineers, research
engineers and applied scientists, data and infrastructure engineers working with robotics
or physical AI, and technical founders exploring embodied intelligence. You do not need to
arrive with a team.

### What to bring

Your laptop, charger, and anything else you rely on for a full day of building.

## The guides

Four walkthroughs, each from zero to a working proof of concept. Full text with commands,
checkpoints, and troubleshooting tables is at https://www.buildspace.tv/.

1. **[LeRobot on macOS](https://www.buildspace.tv/#g1)** — first-time SO-101 setup on Apple Silicon. The
   published LeRobot instructions assume Windows or Ubuntu, so port names, permissions, and
   GPU advice differ. About 45 minutes, most of it calibration. macOS uses
   `/dev/tty.usbmodemXXXXXXXX`, not `/dev/ttyACM0`; there is no CUDA wheel for Mac and
   none is needed (MPS works); terminal camera access must be granted in Privacy and Security.
2. **[Sim teleoperation](https://www.buildspace.tv/#g2)** — hold a real leader arm and watch a physics-simulated
   SO-101 mirror it on a cloud GPU, recording demonstration episodes scored the moment you
   stop. Needs two terminals: the sim session and the bridge.
3. **[Policy training](https://www.buildspace.tv/#g3)** — generate episodes in simulation, train an ACT or
   Diffusion policy, evaluate it closed-loop, and run it on a real arm. About 1.5 hours,
   most of it unattended. Train on commanded joint targets, never measured positions.
4. **[Unitree on Antioch](https://www.buildspace.tv/#g4)** — spawn a quadruped or humanoid and verify it on
   physics rather than video. Isaac Sim and Isaac Lab expose completely different Python
   entry points for the same robot; pick one engine per project.

## Teams and projects

Teams and project submissions are public data, available as JSON:

- Teams: https://www.buildspace.tv/api/teams.json (1 registered, 1 looking for members)
- Projects: https://www.buildspace.tv/api/projects.json (17 submitted)
- Full API: https://www.buildspace.tv/openapi.json

## Links

- Register: https://luma.com/nkknxvrz
- Discord: https://discord.com/invite/nN58zxSTFR
- Source: https://github.com/buildspacetv/sprint