From video to
robot intelligence.

We develop AI software for physical AI, built on two cores: training data generated from video, and safety prediction with a world model.

SIMULATION · Isaac Lab
Autonomous control trained on Anchor data

The bottleneck is training data.

Teaching a robot a new task requires a large amount of training data. Preparing it takes months and costs on the scale of tens of millions of yen, which makes it the biggest bottleneck in robot deployment.

Cost and time

Every site and task needs a dedicated team and heavy GPU spending to build the data.

Mocap does not scale

A dedicated suit, a studio, one person per session. It cannot keep up with the number of sites.

The Sim2Real gap

It works in simulation, then breaks on the real robot because of real-world variation.

HCS Anchor

Turn video into training data that runs on real robots.

Reconstructs both the person and the scene in 3D, then retargets to your robot. The output is checked against physical constraints and goes straight into RL / IL.

Input

  • Site cameras and work logs
  • Phone footage and public video
  • Dashcam, warehouse, factory logs

HCS Anchor

  • Separate people, objects, scene
  • Reconstruct 3D pose
  • Extract contacts and timing
  • Retarget to the robot
  • Score feasibility

Output

  • Joint trajectories, contacts, scene
  • Physically validated
  • Straight into RL / IL
Months → weeks
New task bring-up
New sites need incremental training only.
Zero equipment
No motion capture
Start with the video you already have.
90%+ accuracy
Mocap-level reproducibility
Feedback from the real robot raises accuracy every round.
Person and object segmentation01Segment
3D pose reconstruction023D pose
Contact detection03Contact
Retargeted to a Unitree G104Retarget

Actual output from a single handheld clip (Unitree G1)

Accuracy, proven on a public benchmark.

SLOPER4D (a public CVPR dataset) is to motion estimation what ImageNet is to image recognition and MMLU is to LLMs. Measured under the VideoMimic (CoRL 2025) evaluation protocol, Anchor cuts human motion reconstruction error by 34% versus WHAM and 17% versus TRAM, and it is the only one that also reconstructs floors, steps and obstacles.

WHAMCVPR 2024 · Max Planck
18.9cm
TRAMECCV 2024 · UPenn
14.9cm
HCS Anchorthe only one that also reconstructs the scene
12.4cm

Shorter bars are better (WA-MPJPE). Details and reproduction steps are on the tech page.

HCS Aegis

Five seconds before danger happens.

Predictive safety built on a world model. It catches danger in advance with detection accuracy above the latest research, and avoids it by correcting the trajectory instead of an emergency stop. It attaches to your existing safety devices.

AEGIS · rollout
horizon = 5.0s
simulationFuture prediction by the world model
AEGIS · real
horizon = 5.0s
5 sec ahead
Future state prediction
Catch danger early and avoid it without stopping.
2x
Detection accuracy
Compared with PRISM, the latest research.
78% fewer
False positives
No unnecessary stops, no recovery waits.

Protect uptime

Fewer false alarms and stops, no waiting for recovery.

Protect hardware

Prevent arm and body damage caused by abnormal motion.

Protect people

Prevent harm caused by trajectory deviation.

* Simulation results. Aegis does not replace certified safety devices.

HCS Accord

The fleet that keeps working when the signal drops.

Semi-decentralized fleet control (patent pending). Each robot stays autonomous and connects only when needed. Adding more robots does not break the system.

97%
Task completion
2 to 4 Unitree G1 robots working together. 92% with 8.
40kg
Cooperative carrying
Carried together by 8 robots.
Runs offline
Semi-decentralized
Keeps working where the signal is weak.

* Simulation results on NVIDIA Isaac Gym / Lab.

If you have video, you can start.

Delivered as ROS / ROS2 packages that attach to your existing system.

Humanoids

Refine human work footage into training data and shorten new task bring-up.

Automotive

Turn dashcam footage into development data for active safety and driver assistance.

AMR / AGV

Keep your drive control as is, and add predictive safety as an external module.

Drones

Predictive collision avoidance and fleet control where the signal does not reach.

Contact us

A validation with your own video starts in 2 to 4 weeks, without touching your site. Technical and partnership inquiries are also welcome.

Company

Company nameHumanoid Control Systems Inc.
BusinessDevelopment and provision of AI software for robots
FoundedJune 2026
LocationKita-Aoyama, Minato-ku, Tokyo, Japan
CEORintaro Suzuki