FigureBot Unveils Omni-7: A Home Robot That Learns Your Household Routines in 72 Hours — Without Pre-Programming

TOKYO — FigureBot, a Japanese robotics startup spun out of the University of Tokyo's JSK Robotics Laboratory, has launched Omni-7, its first commercial home robot. At a retail price of $3,999 (or $129/month via a Robotics-as-a-Service subscription), Omni-7 represents what the company claims is a fundamental departure from every home robot that has come before it: it learns your home entirely from scratch, through its own exploration, with zero pre-programmed routines.

The announcement challenges the prevailing approach in consumer robotics, where robots arrive pre-loaded with fixed task libraries (vacuum this, mop that) and require extensive manual configuration to adapt to individual homes.

The Core Innovation: Embodied Curiosity Learning

FigureBot's technology is built on Embodied Curiosity Learning (ECL), a reinforcement learning framework developed by the company's academic founders over seven years of research. Unlike traditional robotic task learning — which requires a human to demonstrate each task or write explicit reward functions — ECL enables the robot to develop useful behaviors through unsupervised exploration driven by an intrinsic curiosity signal.

How ECL Works in Practice

When Omni-7 is first powered on in a new home, it begins an autonomous exploration phase lasting up to 72 hours. During this period:

1. Curiosity-Driven Navigation: Omni-7 moves through the home using an on-board 4D lidar (Livox Horizon 2) and stereo RGB depth cameras, building a semantic 3D map. Rather than navigating randomly, the robot's navigation policy is weighted toward unexplored areas and novel objects — areas that are visually or spatially uncertain generate a higher curiosity reward.

2. Object Interaction Learning: When Omni-7 encounters an object, it performs a brief manipulation sequence (push, pull, grasp, tilt) to observe the outcome. This data is used to build an object affordance model — an internal understanding of what each object is "for" based on how it behaves in the physical world.

3. Sparse Reward Discovery: Human-defined reward functions (e.g., "kitchen should be clean") are replaced by inverse reinforcement learning (IRL) from observed human behavior. If Omni-7 observes a human performing a task — loading a dishwasher, folding laundry — it infers the underlying goal and attempts to replicate the behavior.

4. 3D Semantic Memory: All learned routines are stored in a 3D semantic memory graph — a spatial database that links objects, locations, times of day, and associated actions. The robot knows that "coffee mug + left cabinet + 7:30 AM = morning routine" and "toy blocks + floor + 4:00 PM = play area."

The Result

After 72 hours of autonomous exploration, Omni-7 produces a Household Routine Model — a compressed, structured representation of the home's patterns, preferences, and recurring tasks. The robot can then execute these routines proactively, without prompting.

The company reports that in beta testing with 180 households across Japan, the UK, and the United States:

• Omni-7 correctly identified 94% of recurring household routines within 72 hours
• The robot performed an average of 23 proactive tasks per day by day 4 of deployment
• User satisfaction scores (Likert 1-7) averaged 5.9, compared to 4.1 for the best competing home robot (Unitree H1)

Hardware Specifications

Omni-7 is a whole-body humanoid robot standing 165cm tall, with:

• 61 degrees of freedom (11 per arm, 7 per leg, 6 in the torso, 5 in the neck, 6 in each hand)
• Whole-body torque control for safe physical interaction in proximity to children and pets
• Livox Horizon 2 4D lidar + 6x RGBD cameras for 360-degree environmental sensing
• NVIDIA Jetson Thor compute module (1,400 TOPS dedicated AI compute)
• 2.4kWh solid-state battery providing 8 hours of active operation
• Dual-arm payload capacity of 5kg per arm

Safety Architecture

FigureBot has implemented a triple-redundant safety architecture:

1. Hardware-level torque limiting on all joints (force capped at 15N for contact with humans)
2. Reactive reflex layer — a dedicated micro-controller processes contact sensors at 10kHz, overriding higher-level planning within 2ms of contact detection
3. LLM-based intent modeling — a lightweight language model (3B parameters) runs on-device to interpret human verbal commands and predict human movement trajectories to avoid collisions

Omni-7 has been certified under ISO 13482 (Personal Care Robot Safety) and IEC 61508 (Functional Safety) at SIL 2 level.

Software Platform: FigureBot Home OS

All robot intelligence runs on FigureBot Home OS, a custom robot operating system built on top of ROS 2 Humble. The OS supports:

• Open skill marketplace: third-party developers can publish new skills (similar to smartphone apps), subject to FigureBot's safety review
• Multi-robot coordination: up to 8 Omni-7 units can share learned routines within a single household
• Privacy dashboard: users can view, export, or delete all data the robot has collected, with on-device processing ensuring no raw video leaves the home

Pricing and Availability

Omni-7 is available for order today in Japan, the United States, the United Kingdom, Germany, and France, with shipping to begin in January 2028. The $3,999 outright purchase includes:

• Omni-7 robot
• Docking station
• 1 year of FigureBot Home OS updates
• 24/7 remote support

The RaaS subscription ($129/month, 36-month minimum) includes all hardware, software, and proactive maintenance including annual servicing visits.

Funding

FigureBot has raised $480 million across four funding rounds, led by SoftBank Vision Fund 3, with participation from Toyota Ventures, Samsung Next, and the Japanese government's New Energy and Industrial Technology Development Organization (NEDO). The company employs 890 people across Tokyo, San Francisco, and Munich.