AstraYams
AstraYams aims to make teaching robots new jobs faster and cheaper by generating the training data they need.
Core trick
Generates MuJoCo task packages with CAD meshes and reward code, then produces varied demonstrations for robot training.
Description
AstraYams aims to make teaching robots new jobs faster and cheaper by generating the training data they need. Describe a task, and Astra builds a simulated world, develops the skill, and generates varied demonstrations for robot models to learn from.
README excerpt
# AstraFactory **An AI that creates the robot skills it’s missing.** Give a robot a task. Astra analyzes its failures, creates targeted simulation and training experience, trains a skill, and evaluates whether it improved. ## First experiment A YAM robot grasps a free custom keyed peg through finger contact, lifts it and attempts insertion into a matching socket. A shared MuJoCo/Gym core loads task-specific CAD, reset rules, reward code and acceptance checks. The peg is a separate dynamic body, with no attachment to the gripper. ## Repository - `assets/robots/yam/`: official I2RT YAM URDF, meshes, upstream license and pinned source manifest. - `docs/ENVIRONMENT.md`: task-package contract, running instructions and limitations. - `tasks/keyed_insertion/`: generated CAD meshes, physics scene, reward and evaluator. - `tasks/yam_keyed_insertion/`: articulated YAM task, physical joint control and insertion acceptance contract. - `tasks/yam_contact_insertion/`: active task with moving fingers, free peg and contact-only manipulation. - `src/astrafactory/`: reusable Gym core and camera observations. - `tools/`: CAD generation, environment checks, rollout recording and PPO training. ## Status Four small learned controllers (3,242 parameters total) now execute contact grasp → lift → alignment → insertion continuously on the official YAM assets. The frozen chain passed **20/20 reserved nominal tests and 20/20 reserved stress tests**. It uses privileged simulator state, hand-desig