SIM2REAL CHALLENGE
Data Center Cabling Challenge
Train a robot to clean, connect, and swap data center fiber. Prove it in simulation. Take it to the real bench.
Submissions open. Download the sim and start building.
SUBMISSION DEADLINES
Build toward the real bench.
WHY THIS MATTERS
Small connections.
Real consequences.
AI data centers have thousands of high speed cables. Cleaning, plugging, and unplugging them is slow, manual, and error prone. A dirty fiber end or a half seated latch can take a GPU cluster down.
This is a real task for a major AI cloud provider. Two robot arms face a rack of real ports. Your learned, vision-based policy must make the connection, seat the latch, and remove it safely.
THE TASK
Four levels. Every connection counts.
Learn the connection.
Then handle the whole sequence.
LC fiber
Clean. Connect. Confirm.
Clean the fiber end. Plug it into a dense LC patch panel. Confirm the link light, then remove it.
MPO fiber
Get the orientation right.
Clean the fiber end. Align the key and connect it to a transceiver. Handle both push-pull and push-only connectors.
Transceiver swap
Out with the old. In with the right one.
Unplug the fiber. Remove the transceiver and seat a new one. Verify it is the right module, then reconnect the fiber.
Ordered run
Eight cables. The right order.
Start with eight labeled cables hanging above the bench. Connect each to its assigned port, in order.
Warm up: insert and remove an RJ45 copper plug. Not scored.
SIM2REAL
Train in sim. Prove it on hardware.
The NVIDIA Isaac Lab development simulator is available now, with all four task scenes, training tools, and a policy template. Hardware calibration and full task qualification are still in progress.
01 / Train
Build a learned policy using camera images. Train and validate in the sim.
02 / Qualify
Score in simulation on the public leaderboard. Top teams earn time on the real bench.
03 / Transfer
Run your policy on hardware using the provided adapter. The real bench decides the winner.
A proof of concept with a major AI cloud providerFor the team with the highest real bench score.
HOW YOU WIN
Right the first time.
Then fast.
- First pass yield comes first. How many inserts seat and light up on the first try, without a retry. Speed breaks the tie.
- The bar: beat an average human technician on first pass yield, at the same speed or faster.
- No damage. Any damage to a port, fiber, or transceiver scores zero for that run.
Inside the real bench
- 2× Universal Robots UR5e class arms: six axes, force sensing.
- Parallel two finger grippers with swappable 3D printed fingertips.
- 2× wrist RGB-D cameras and 1× overhead RGB-D camera.
- Standard 19 inch rack with a 2U OSFP switch face (SN5600 class, 64 ports) and a 1U RJ45/SFP switch.
- RJ45 Cat6, LC-LC and MPO fiber cables, plus OSFP and SFP transceivers.
- E-stop and low force limits. Humans stay outside the arm workspace during runs.
START BUILDING
Your sim. Your policy. Your first run.
Download directly.
No repository invitation needed.
Development snapshot b31c816 · 301 KB · All four levels included.
01 / Set up
Linux x86_64, RTX GPU with 24 GB VRAM, and 100 GB free disk. Unzip, initialize the local Git snapshot, and follow the included README to install the pinned Isaac Lab runtime.
02 / Train and test
Run Level 1. Train from RGB-D images, joint states, and force readings. Keep the supplied 14-action policy interface and check your container before evaluation.
03 / Submit
Upload a ZIP with policy code, Dockerfile, weights or immutable model reference, and a short README. Add a real MP4 from the sim. ZIP and video: 100 MB combined; MP4: 50 MB maximum.
Policy interface and simulator status
Implement reset(obs) and act(obs). Observations include three 640×480 RGB-D cameras, joint position and velocity, gripper openings, wrist force/torque, target port and elapsed time. Return six motion values and one gripper command for each arm, normalized to [−1, 1].
This is an uncalibrated development candidate. Full Level 2–4 demonstrations and final-source qualification remain unfinished. Uploaded entries are reviewed separately from official leaderboard scoring. Normal RunPod Pods support trusted training and playback; container evaluation needs a qualified Docker-capable RTX host.
Simulator ZIP SHA-256:
a6dec5756ef9c742eb2138c6fc1cb7a1624250b7532ff3718345b68204a605b6What we need from a submission
Submit a learned vision policy, a short writeup, and real simulation footage.
- A learned policy that uses RGB-D camera images to produce arm and gripper commands in the sim. Joint states and force readings may supplement vision.
- Use the supplied container template and its reset/act interface. Return 14 normalized arm and gripper commands from the provided observations. Include reproducible packaging instructions.
- Must run on the real bench with no code changes beyond the provided hardware adapter.
- Short writeup: approach, model, training data, compute used, and how the learned policy uses vision.
Rules
- Any learned model or training method. The policy must use camera images to perform the task.
- Scripted task policies and fixed action sequences are not eligible. Organizer diagnostic scripts are not valid submissions.
- Use only the provided policy observations. No ground-truth object poses or hidden simulator task state.
- Default: results and policies are open source after the challenge. Teams who need a different arrangement contact us.
- Download the sim without registering. Sign in to submit your policy. Published entries and project ZIPs are public.
- HIM Arena may feature results, videos, and team names on the site and in marketing.
PUT YOUR POLICY TO THE TEST
Submit your policy
Upload your project ZIP and a real simulation video. Sign in to submit. Submission deadlines: Level 1 on November 15, 2026. All four levels by December 31, 2026.
Your upload will be connected to your builder profile.
Download the sim and instructions above. Sign in when you are ready to submit.
Sign in