Notes from the lab

The journal.

Updates, deep-dives, and behind-the-scenes notes from the team building the first robot for the home.

← Back home
Journal··TY Lok

Why We're Not Building Humanoids

Our Path to Useful, Affordable AI Robotics

Humanoid robots are having a moment. Every demo reel shows bipedal machines jogging, dancing, or climbing stairs. But here’s the truth behind those dazzling visuals: These highly edited, tightly controlled demos show the absolute best moments of a system still wrestling with fundamental challenges.

While the vision of a robot helper that looks and moves like us is compelling, the complex human form presents unnecessary, engineering challenges that make these robots prohibitively expensive and complicated for everyday home use.

 

Legs: A Safety Liability

  • Balance is Power-Hungry: Standing upright requires constant, active balancing, leading to continuous computational and power drain. Bipedal robots are always on the verge of falling. Without the need for complex sensor and high-torque motors, wheeled locomotion are 2x more energy efficient.

  • The Mobility Mismatch: Homes have predictable flat floors. Wheels are simpler to manufacture, maintain, and can carry heavier loads with ease. The only drawback are stairs.

  • The "Fall" Factor: A wheeled or track-based robot can't really fall; a bipedal robot can. Every fall risk expensive damage and hurting kids and pets. 

Hands: Expensive to Build, Expensive to Maintain

  • Unaffordable: A typical humanoid hand has 20 to 30 degrees of freedom (DOF). Each DOF requires an actuator, a gearbox and multiple sensors, driving up costs.

  • Inefficient: Analysis shows that most required manipulation tasks in homes can be handled with task-specific end-effectors requiring a maximum of 4 DOFs.

  • Monthly Maintenance: Repetitive motions cause high wear and tear, Xpeng noted their humanoid robot's hand require monthly replacement.
     

Intelligence: Training for the Unnecessary


  • Training Complexity: Teaching an AI to precisely control dozens of DOF in a human-like hand is incredibly complex. For a simpler robot, the AI only needs to solve a simpler problem.

  • Function, Not Form: The point of a domestic robot is for it to do your chores. The AI doesn't need to learn how a human would walk or do laundry. It simply needs the most efficient path and method to get the job done. AI is far more effective when its job is clearly defined, and its physical platform optimized.

 

No Running Before We Learn to Walk

There may be a day where highly generalist robots does indeed populate homes and industries. But engineering reality demands pragmatism. Today, and for the foreseeable future, these highly complex, multi-DOF humanoid robots are simply too expensive, fragile, and inefficient for household use.

At Sapir Robotics, our mission is to deliver value and reliability now. We ditch the complexity for optimized forms like our stair-climbing wheeled mobility system and low-DOF grippers. 

The future is cool, but in the meantime, we’re making real products for real homes.

 

Company··Sapir Team

Home Robot Report Card: Ranking the Best Releases of 2025

At Sapir Robotics, we track the market closely to identify solutions that actually deliver value in a real-world home environment. Here is a quick recap of the most interesting 2025 launches and updates and what we think each one got right.

1. The Industrial Workhorses

Focus: Solving the problems of safety, durability, and non-stop operational integrity needed for real-world deployment in structured environments. This group proves the technology is ready to work.

• UBTECH "Walker S2" (Continuous Uptime) Autonomous battery swap and 24/7 deployment prove reliable industrial viability.

• Agility "Digit 2025" (Regulatory Compliance) OSHA-audited and deployment in structured industrial settings, safely integratable into human workflows.

2. The Consumer AI and Data Labs

Focus: Solving the "intelligence gap" through novel training methods, superior perception, and strategies to efficiently acquire and generalize complex skills in unstructured environments like the consumer space.

• Sunday Robotics "Memo" (Data Acquisition) Data Capture Gloves system for scalable training data proves data strategy beats hardware complexity.

• 1X "NEO" (Remote Tele-Operation) Relies on remote human operators for tasks and training data, turning your home into a training group for a privacy-sensitive R&D program.

• Figure "03" (Production Scalability) Trialed in a BMW car line running live tasks, proves manufacturing is scalable and achievable with high-volume, low-cost.

3. The Hardware Maximalist Testbeds

Focus: Pushing the boundaries of physical capability in terms of high Degrees of Freedom (DoF), power density, and complex mechanism. Often ignoring cost for the sake of long-term component development.

• XPENG "Iron" (All-out R&D) Engineering showcase with max articulation and advanced muscle/battery tech. Pushing the boundaries of physical capabilities, ignoring costs and deployment.

• EngineAI "T800" (Component stress-testing) Extreme performance events like robot fighting to push components beyond normal operating limits. Use spectacle as a marketing tool while quieting piloting retail "cyber staff" in stores.

• Tesla "Optimus Gen 2.5/3" (Hand Dexterity) Tendon-rich hand with massive dataset from vehicle fleet, proves biggest obstacle to fine manipulation are the complex hands.

4. Democratization and Ecosystems

Focus: Reducing the barriers to entry (cost or complexity) to rapidly expand the developer base and accelerate the creation of new robotic skills via the community.

• Unitree "R1" (Affordability) Low cost (starting at $6k) bipedal robot, opening developer/mass market access for rapid, crowd-sourced testing.

How the Specs Stack Up

RobotTotal DoFDoF / HandAI ComplexityData Capture StrategyNotes
Walker S25211Medium – BrainNet 2.0 Cloud swarm co-agentDual stereo camera Force-torque sensorsAutonomous battery swap for 24/7 operation
Digit (2025)28N/ALow – Agility Arc Visuomotor controller LLM Cloud fleet managementLiDAR + depth cameras On-site real-world dataDeployed in warehouse logistic settings; OSHA-audited
Memo274Medium – ACT-1 Zero-shot learning VLASkill Capture Glove records human motion (10M Eps.)Wheeled based; sacrifices stairs for efficiency
NEO7522High – Redwood AI 160M param VLA Built-in LLM 2070 TFLOPSDual stereo cameras Remote teleop$20k preorder; need to schedule human operators for tasks
033010High – Helix 7B param OS VLM 8M param visuomotor controllerVideo + Sim Teleop imitation VLM adds hindsight instructions from onboard camera
Iron8222High 3000 TOPS VLT/A/MMulti-sensor fusion Imitation learningSolid state battery, bionic muscles, can customize robot body shape
T800437High Real-time perception 2000 TOPSMulti-Sensor Fusion (Vision, Tactile, Force) 360° LIDARTactile sensing hands; active cooling system
Optimus (2.5/3)7625High – FSD perception End-to-end learning Pure Vision SystemImitation + Sim Automotive data stackPlanned internal deployment at Tesla factories
R1262Medium – UnifoLM 100 TOPS 8-core CPU + GPUBinocular camera Imitation learningBattery life of 1 hr

Our Takeaways

1. The real fight is hands + data + training. We have yet to see any company perform chores reliably and effectively.

2. There are companies focusing on heavily simplification of body (Sunday) while others are trying to precisely follow human-like gait (XPENG) leading to drastically different estimated prices.

3. Data pipelines are finally differentiated.

4. Durability of hand joints is an issue. XPENG admits complex hands requires replacement every month 13 .

5. Edge processing power is a critical challenge to deployment in homes. Traditional robots often rely on cloud connectivity for complex computation, introducing privacy and latency issues in dynamic domestic applications.

roboticsAIhome-automation
Company··Sapir Team

Welcome to Sapir: Building the Future of Home Robotics

We're excited to share our vision for transforming everyday life through autonomous home robotics. Learn about our mission and what we're building.

At Sapir, we believe that home robots should be as transformative as the washing machine once was. Our mission is to build the first robot that can autonomously vacuum, cook, do laundry, and clean your home.

We're starting with a clear vision: to create technology that genuinely improves daily life. Just as the washing machine freed up countless hours and changed how we think about household chores, we believe autonomous home robots will do the same for the next generation.

Our team brings together expertise in robotics, AI, mechanical engineering, and product design. We're not just building a robot—we're reimagining what's possible in the home.

Stay tuned for updates on our progress. We're just getting started, and we can't wait to share more with you soon.

roboticsAIhome-automationcompany-news