Which robot, and how to build your first prototype

Which robot shape for physical AI?

Not every robot embodiment benefits equally from the 2026 wave of physical AI progress. The more controlled the environment, the more bounded the task, and the more safety is already under control, the faster a vision-language-action model or a diffusion policy pays off. Conversely, every extra degree of freedom, every metre of unstructured terrain and every kilogram of balance to manage pushes the maturity horizon further out. Here is a decision matrix by embodiment type, to help you decide where to invest first.

These categories already have verifiable industrial representatives. Agility Robotics sells its bipedal Digit (175 cm, 65 kg, 28 degrees of freedom) for around $250,000 and deploys it in US warehouses for tote sorting in segregated zones. Boston Dynamics, for its part, has deployed its mobile manipulator Stretch at DHL, Gap, Maersk, Arvato, Otto Group, H&M and NFI, reaching throughput above 500 cartons an hour as it rolled out at Lidl in 2026. On the wheeled-humanoid side, Galbot G1 (173 cm, 85 kg, 47 degrees of freedom, omnidirectional wheeled base) already runs in a Beijing pharmacy, and Unitree has been selling its G1-D, a wheeled variant of the G1, since 2026, priced from $38,800 to $66,300 depending on configuration, built for data collection rather than walking. In inspection, ANYbotics' ANYmal X, certified for explosive zones, runs at Shell, Equinor and BASF, while Boston Dynamics' Spot completed multi-week trials at BP's Whiting refinery.

For a first prototype, start with a fixed arm in a cell, not a humanoid. That combination maximises your odds of shipping a convincing demonstration in a few months rather than several years.

The learning curve, no sugarcoating

For a team of 1 to 2 engineers comfortable with Python but with no robotics experience, here is a realistic schedule, built from field feedback of the LeRobot community rather than a marketing scenario.

Honestly budget 2 to 3 months for a convincing first prototype. The good news: a single weekend is enough for a minimal demonstration. With an SO-101 kit, it takes about two hours to record 50 demonstrations, and training a first ACT policy fits into an afternoon on a decent GPU. It is the iterations that follow, not the first attempt, that consume the two to three months.

The classic walls almost every team hits:

• Demonstration quality: hesitant or inconsistent trajectories teach the policy to hesitate in turn.
• Calibration drift: a camera or arm that shifts by a few millimetres between sessions is enough to break an already trained policy.
• Lighting: a model trained in morning light fails in the afternoon if lighting was not varied during collection.
• Reproducibility: a policy that succeeds nine times out of ten on a given day must be revalidated on another day, under different conditions, before drawing any conclusion.

The concrete kit for your first prototype

Here is a hardware list verified at the time of publication, not a wish list. Prices remain orders of magnitude: check them before buying, this market moves fast.

On the software side, Python and Hugging Face's LeRobot library remain the standard entry point in 2026: actively maintained, and already backing more than 58,000 community datasets published on the Hub, up from 1,145 at the end of 2024. Add ROS 2 if your robot moves, and simulation as an option: MuJoCo, free and open source, or NVIDIA Isaac Lab for massively parallel training of locomotion policies.

Total budget for a discovery-stage build: under $2,000 excluding the PC, with an SO-101 arm, two webcams and optionally a Jetson Orin Nano. That is the price of a good laptop, not an industrial project.

The data you actually need

How many demonstrations does it take for a policy to work? LeRobot's documentation for ACT (Action Chunking Transformer) gives clear, documented orders of magnitude.

Collection happens through teleoperation with the SO-101's leader arm: you move the leader arm by hand, the follower mirrors the motion, and every trajectory is recorded automatically. To get started without collecting everything yourself, the LeRobot Hub and the Open X-Embodiment dataset (more than a million trajectories from 22 robots across 21 institutions) provide a base for training or pretraining.

The golden rule, documented by 2026 research on demonstration selection: quality and diversity matter more than volume. Fifty clean, varied demonstrations outperform two hundred sloppy, repetitive ones.

One strategic point to keep in mind: public data lets you get started, but it is the data from your own processes, collected on your own line, that forms the asset that actually matters. That proprietary collection, far more than the choice of model, is what becomes your lasting advantage.

Constraints and evaluation

A prototype that works on a bench is not yet a safe system, nor a seriously evaluated one. Three disciplines you should not skip.

Safety from the prototype stage: limit speeds, clear an exclusion zone around the arm during trials, and never treat a lab prototype as ready for a human operator standing nearby. Moving to production then falls under the EU AI Act and the associated machinery standards: see our EU AI Act and standards pillar for up-to-date deadlines and responsibilities.

Evaluating properly means measuring a success rate over a meaningful number of consecutive trials, with variation in position, lighting and objects, rather than keeping the best filmed take for a demo. A policy that succeeds nine times out of ten across varied trials is a solid result; a single successful video is not.

Moving from prototype to industrial pilot requires a different set of requirements: reliability, maintenance, integration with existing systems, team training. Our factory director's roadmap details the pass criteria and a realistic timeline.

When to stop: if the required cycle time is under 2 seconds or the required precision falls below a millimetre, classic automation (SCARA, delta, deterministic industrial vision) remains the right answer. Physical AI excels at variability and uncertainty, not at perfect repetition.