Training guides for every model and arm
Pick a policy and a robot arm and get the exact training guide for that pair: GPU tier, trainer defaults, dataset format and the run itself, step by step.
ay-robots.comCrawled Nov 16, 2026
Pick a policy and a robot arm and get the exact training guide for that pair: GPU tier, trainer defaults, dataset format and the run itself, step by step.
Every ay-robots CLI command, verified against the code: install, pair, devices, calibrate, record, datasets, run and mcp, with the options each takes.
Copy-paste MCP setup for Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Cline, Zed, Gemini CLI, ChatGPT developer mode, Grok and the OpenAI Agents SDK. One hosted endpoint, public tools without a key.
The AY-Robots desktop client allows users to record imitation learning datasets with an SO-100 robot arm without Python setup, also available as a CLI.
A physical SO-100 arm is online and anyone can control it from the browser. No account, no installation, no simulation. One visitor drives at a time, the rest queue.
Users can test robot learning models before purchasing hardware by driving an SO-100 in the browser, comparing models on benchmarks, or fine-tuning on a rented GPU.
The page details how credit is earned on the platform, including earning it instead of paying for it, and clarifies that training and inference use rented cloud GPUs billed per minute, with datasets sourced from public repositories or the marketplace.
Record fifty usable demonstrations with an SO-100 robot arm, review them, and upload a LeRobot dataset ready for training.
Deploy a trained policy to a real SO-100, evaluate it honestly, diagnose the common failure modes, and decide what data to record next.
Assemble, wire and calibrate an SO-100 robot arm, install the AY-Robots client, and move the arm for the first time. No prior robotics experience required.
Four hands-on guides for the SO-100 robot arm: setup and calibration, recording an imitation learning dataset, training a policy, and running it on real hardware.
Users can start a training run on an uploaded SO-100 dataset, select a policy architecture, monitor the loss curve, and choose the appropriate checkpoint.
Simple, transparent pricing for robot teleoperation. Pay per hour with no hidden fees. Rates from €8-22/hour based on Remote Data Collector region.
AY-Robots offers pay-as-you-go and Starter plans with monthly invoices, outlines operator earnings via Stripe Connect, and explains refund and dispute processes.
ACT and SmolVLA both train on 24 GB for 1 to 3 USD. One is 80 million parameters from scratch, the other 450 million pretrained with language.
Six head to head comparisons of the policies this platform trains, with the GPU tier, the trainer defaults and the inference latency each one really has.
GR00T N1.7 and Pi0.5 performance is compared for SO-100 tasks, with metrics like action step speed (152 ms vs. 485 ms), batch size, and seed dependency.
GR00T N1.7 needs an 80 GB card at 4 to 12 USD a run, SmolVLA a 24 GB card at 1 to 3. Which one your dataset size justifies, and when the upgrade pays.
Pi0.5 costs four times more per run and twice the latency of SmolVLA. The failure pattern that justifies the upgrade, and the ones that do not.
The page lists 20 robot arms from 14 brands, including SO-100, SO-101, Franka, UR, KUKA, and UFACTORY, detailing reach, payload, price, licensing, and LeRobot support, with 14 available as 3D models.
Users post details of missing robot data—arm, task, scope, and hourly rate—and several Remote Data Collectors record it in parallel. The data is compiled into a single downloadable collection, with payment per hour of recorded material and no payouts until acceptance.
Users describe needed robot data—arm, task, episode count, budget, and rights—without creating an account. The draft remains on their device until submission, and the request goes live once budgeted.
A servo scan finds four of six IDs. How a daisy chain fails, why the last responding ID locates the bad cable, plus duplicate IDs and baud rate mismatches.
Cosmos Policy by NVIDIA: 2 B parameters, Code Apache-2.0 weights, released January 2026.
The five policy models this platform trains: GR00T N1.7 and N1.5, Pi0.5, SmolVLA and ACT. Parameters, GPU tier, inference latency and when to pick which.