Station-side toolkit for collect → (optional) train → infer on single-arm setups. I use it mainly on an xArm7 cell; the same patterns cover AgileX-style stations where the arm, wrist cameras, and contact sensors must stay time-aligned.

Code stays private. Below is the public shape of the system.

Scope

Stage	Role
Collect	Pedal-driven episodes: arm state, wrist cameras, depth camera, dual tactile, fingertip force
Teleop	VR Cartesian servo + gripper, recorded in the same episode layout
Infer	Shared runtime hooks so a trained policy can reuse the same device stack

Output is episode folders with synchronized pickles (arm / gripper / cameras / tactile), compatible with our internal training format.

Collection modes

Teach (drag). Pedal starts recording; the arm is put into teach mode; the operator moves the arm and gripper by hand; pedal stops, homes, and writes the episode.

VR teleop. A dedicated teleop process owns arm servo and gripper commands. The collector only records sensors and merges arm/gripper streams over IPC so teach-mode and teleop-mode land in the same schema.

Both paths share session lifecycle: bring-up → record → flush cameras → validate → save → home.

Sensors (logical)

Without naming lab IPs or vendor debug steps:

• Wrist RGB from an onboard computer (decoded live, flushed at episode end)
• Scene depth camera
• Dual tactile streams (GPU inference path matters for rate; CPU fallback is too slow for 30 Hz)
• Optional fingertip force via serial
• Arm joint + TCP state; gripper feedback

A USB pedal drives record / save / quit so the operator never leaves the workspace to hit a keyboard.

Design choices that mattered

• One episode schema for teach and teleop. Training code should not care how the human drove the arm.
• Flush on stop. Wrist video buffered over the network needs an explicit end-of-episode drain or the last second disappears.
• Session restart cadence. Long-lived collector processes leak; we force a clean relaunch every N episodes.
• Validate before keep. Broken episodes (empty streams, length mismatch) fail early instead of poisoning a dataset.

Limits

This is lab infrastructure, not a released product. Hardware bring-up, calibration, and model training live in other repos. Publicly I can only describe interfaces and responsibilities — not configs, network layout, or source.

If you work on contact-rich manipulation, the useful idea is simply: treat collection as a first-class runtime, with the same contracts you will need at inference time.

---

← Back to projects