Purpose Robotics — Semi-Humanoid Robot Under $2,000

Purpose Robotics set out to answer a simple question: how capable a semi-humanoid manipulator can you build for under $2,000? The result is a torso-and-head platform with dual arms, each with enough degrees of freedom to reach across a tabletop workspace and manipulate everyday objects.

The chassis, head, and arm links are 3D-printed, keeping structural cost low while allowing rapid iteration on link lengths and joint layouts. Each arm is a chain of servo actuators terminating in a custom gripper, sized to pick up small consumer items reliably.

Alongside the physical robot, I built a MuJoCo simulation to iterate on manipulation behavior without hardware in the loop for every change. The chassis, telescopic lift, torso, and head come straight from the robot's Onshape CAD; the arms come from their own URDF definitions, with collision geometry generated by decomposing the real meshes into convex pieces rather than approximating them with primitives.

A scripted controller drives both arms through a bimanual fold on a simulated T-shirt, using MuJoCo's flex cloth solver for the garment physics. The same environment is wrapped as a Gymnasium task with domain-randomized actuator offsets, set up to train a residual reinforcement-learning policy on top of the scripted motion. Joint trajectories export in the hardware controller's own units and rate limits, so a validated simulated motion is a direct candidate for hardware replay.

The robot demonstrates autonomous pick-and-place, identifying objects on the table and placing them into a bin without manual teleoperation for each pick — showing that a functional humanoid-style manipulator is achievable well outside typical research-lab budgets.

In simulation, the scripted bimanual fold reliably brings the shirt's hem up onto its shoulder seam and reduces its footprint by roughly a quarter each run. The residual-learning setup is in place and trains against that scripted baseline, but has not yet been shown to beat it — it's an active area of the project rather than a finished result.