Work samples for robotics expert and AI data annotation roles: runnable simulations with logged failures, plus a real annotation guideline with worked examples.
Rea Ara — five years as Senior TPM at Google, NYU M.S. Cybersecurity (in progress). Contact: rea@zetra.dev
simulations/reacher_mujoco.py builds a 2-link planar arm in MuJoCo
(MJCF defined inline, no external assets), samples random targets in the
workspace, solves 2-link planar IK analytically, and drives the joints
with PD control.
python3 reacher_mujoco.py --trials 30 --out reach_log.csv
Latest run: 24/30 successful, mean final error 1.0 mm on successes.
simulations/reach_log.csv is the actual log from that run, with a
failure taxonomy: unreachable (target outside the arm's workspace),
timeout (no convergence in the step budget).
simulations/pick_place_pybullet.py runs a KUKA iiwa through a full
pick-and-place: approach, descend, grasp, lift, transport, release,
retreat, with randomized cube starts and per-trial failure logging
(grasp_miss, drop, timeout). Needs pip install pybullet
(a local build issue in this sandbox kept it from executing here; the
script is complete and runnable anywhere PyBullet installs).
data-annotation/labeling-guideline.md— a full labeling guideline for robot manipulation episodes (LeRobot / Open X-Embodiment / DROID style): episode fields, temporal segment definitions with boundary rules, outcome and failure taxonomy, per-frame fields, quality rules.data-annotation/sample_annotations.json— three episodes annotated in that schema: a clean success, a grasp_miss with recovery ending in success, and a drop ending in failure.
- MuJoCo, PyBullet, and Isaac Sim are all free; these samples use the first two. Next: the same reaching task on a simulated SO-100 arm via the LeRobot datasets.
- Everything here is original work written for this portfolio.