I build robots that learn to
see, move, and act.

Industrial pick & place · XP Robotics

Vision-guided carton picking on a HITBOT 6-axis arm

A 6-axis HITBOT arm picks up a carton whose position it was never taught. No fixtures, no hardcoded coordinates, no ROS — the entire stack is Python.

1. An Intel RealSense D415 captures a single aligned depth frame, deprojected into a point cloud, voxel-downsampled, and fitted for the support plane.
2. Anything standing a few centimetres above that plane becomes a candidate. DBSCAN clusters the points; each cluster gets an oriented top face and a bounding size.
3. The planner checks whether the carton's top surface is large enough for the four-pad suction to seal. If it isn't, the arm doesn't move toward it. Of the rest, it takes the nearest.
4. The pose is transformed through hand-eye calibration into base coordinates. The pick is three Cartesian moves — hover, contact, lift — each run as interpolated waypoints with an IK solution at every step, aborting if the solver flips configuration mid-path.
5. Descent intentionally overshoots the expected contact by ~13 mm so the suction bellows compress; at 3 mm the pads were only barely touching the lid. Then: vacuum on, dwell, lift, swing the base joint, descend, release, home.

The carton is placed at a PLC-controlled station — the first step toward a real industrial deployment.

• Python
• RealSense D415
• Point-cloud segmentation
• DBSCAN
• Hand–eye calibration
• Cartesian IK
• 4-pad suction
• PLC station

Bimanual manipulation · XP Robotics

Two-arm autonomous cloth folding

Two LeRobot arms fold a cloth in half together, trained on just 10 clean teleop demonstrations.

I recorded three camera streams — two wrist cameras and a front camera to locate the cloth — and trained a SmolVLA policy. Despite the tiny dataset, the policy generalised well across starting positions of the cloth.

• SmolVLA
• Bimanual
• 10 demos
• 2 wrist + 1 front cam
• Position generalisation

Personal project · RL locomotion

A 14-DOF biped that learned to walk

Mechanical design inspired by OpenDuck. I trained a PPO policy with RSL-RL in MuJoCo, with domain randomization so the behaviour has a real chance of transferring to hardware.

• Actuators14
• Observation space48-dim
• Action space14-dim
• Policy network191,118 params
• Control rate50 Hz
• Deploy targetJetson Orin Nano

The trained policy is exported to a lightweight NumPy implementation for on-device inference. It walks reliably in simulation; next is sim-to-real — calibrating the robot, validating the observation vector, and seeing how much of the behaviour survives outside the simulator.

In parallel I'm building a local AI stack for the robot: an on-board LLM for high-level planning and natural-language interaction, backed by a RAG memory system so it can carry context between conversations.

• MuJoCo
• PPO
• RSL-RL
• Domain randomization
• NumPy policy export
• Jetson Orin Nano
• Local LLM + RAG

Hardware bring-up · XP Robotics

UFACTORY xArm 7 — motion & perception

Brought up a 7-DOF UFACTORY xArm 7: basic joint and Cartesian motion, planned and executed through MoveIt in RViz, with an Intel RealSense camera mounted on the end effector for eye-in-hand perception.

• xArm 7 (7-DOF)
• MoveIt
• RViz
• Eye-in-hand RealSense