TECHNICAL ARTICLES
Implementation decisions, experiments and unresolved problems in robotics. Explore our simulation research by topic.
Project scope and results →
· Simulation research
A 90 m orchard route, a crossing person and two crate layouts: TeamZ connects camera AI, LiDAR, IMU and synthetic GNSS in Isaac Sim. Watch the 78-second independent demo and see what was measured.
Read article →
· Simulation research
A self-funded rover demo connects depth-based obstacle detection, path planning, driving and instrument deployment. Watch the 75-second simulation and compare routes under two rock layouts.
Read article →
· Simulation research
Go2 uses image-based localization to approach a gauge, withholds an unreliable reading, and moves to a configured viewpoint to retry. Video, before-and-after images, measured results and the remaining limits of this simulation.
Read article →
· Simulation research
TeamZ tested microwave opening with Flex-π in two simulation environments. We explain checkpoint–task compatibility, robot differences, success criteria and video timing; this is not a controlled performance comparison.
Read article →
· Simulation research
Harvesting and retaining a mandarin have not been achieved. Diagrams and recorded simulation traces explain virtual tactile feedback with the Unitree Dex3-1 hand, force-triggered stopping and remaining problems.
Read article →
· Simulation research
We refine a 3D model of a mandarin tree, import it into Isaac Sim, and test one fruit's attachment, release, fall and ground contact. G1 harvesting and policy training remain unfinished.
Read article →
· Simulation research
We test G1 across four corridor widths, three entry positions and three headings in 36 Isaac Sim trials. A 5 cm entry shift changes contact at the same 80 cm width. See the complete results and evaluation limits.
Read article →
· Simulation research
TeamZ trained G1 in Isaac Sim and tested 20-second runs over three rough terrains, covering about 39–40 m. We report flight-phase checks, reward changes and failed low-speed trials.
Read article →
· Simulation research
We retarget human motion to G1 and prototype bowing and guiding with model-based whole-body control. See the adjustments, physics-test scope and untested hardware conditions.
Read article →
· Simulation research
We test Play2Perfect with KUKA iiwa and a Sharpa hand: insertion and withdrawal passed 3/3 fixed and 10/10 separate-seed trials. The part starts in hand; clearance is not measured accuracy.
Read article →
· Simulation research
A 1 kg G1 transport test passed at effective friction 0.75; four conditions at 0.4 or below failed to lift. We explain combine modes, floor-contact changes and the one-trial-per-condition limit.
Read article →
· Simulation research
The public HAIC policy carried a 1 kg box in 3 fixed and 10 position-variation trials. We checked placement, hand release and three seconds of rest, without additional TeamZ training.
Read article →
· Simulation research
We connect fixed-camera tracking to a G1 using Visual SLAM, linking person IDs to ROS 2 stop reasons. The article covers timestamps, coordinates and rosbag2 replay in a short single-person test.
Read article →
· Simulation research
Two 30-second G1 trials averaged 3.23 m/s. Stereo and IMU pose output continued, but endpoint error reached 11.25 m. We report the offline evaluation conditions and unresolved limits.
Read article →
· Simulation research
We compare two-handed G1 valve motion under 0–0.6 Nm opposing loads. Rotation and hold checks accompany contact-speed and posture measurements; slip and upright-posture issues remain.
Read article →
· Simulation research
We compare three walking and three running trials in Isaac Sim, plus initialization tests. Results include meter-scale running drift and failed motion conditions, using offline stereo/IMU evaluation.
Read article →
· Simulation research
We integrate the public Chasing Autonomy policy, adjust initial phase after falls, and evaluate 60 seconds of warehouse running. Foot-contact, speed and ground-truth heading correction are documented.
Read article →
· Simulation research
We connect Visual SLAM and Nav2 in Isaac Sim. Extra observations repair an incomplete map; one final run detours around shelving and stops 8.5 cm from the goal. Conditions and limits are documented.
Read article →
· Simulation research
We build maps from G1 stereo images and test saved-map relocalization on another recording. The article distinguishes colored point clouds from visual-feature maps and explains five-point test limits.
Read article →
· Simulation research
We measure G1 leg vibration across four evaluation, model and policy setups. High-frequency variation fell 45.6% in the final setup; one trial per setup does not isolate individual causes.
Read article →
· Simulation research
A B2-W that stands but does not advance motivates a review of rewards, fall termination and leg/wheel observations. This is a design review of public records; original training logs are unavailable.
Read article →LET’S BUILD TOGETHER
Tell us about your task, equipment and current challenges. Through simulation and tests on real robots, we explore feasibility and technical challenges together.
Discuss joint development