Sunday Robotics: Scaling the Home Robot Revolution with Co-Founders Tony Zhao and Cheng Chi

The founders explain why robotics is experiencing its 'GPT moment' - the field has found scalable algorithms (diffusion policy, transformers) but hasn't yet scaled them to consumer products. Classical robotics required custom engineering for each task with no knowledge transfer, but new AI approaches enable generalization across tasks and environments.

• •Robotics has a proven recipe that can scale, but hasn't been scaled to consumer products yet (analogous to GPT vs ChatGPT)
• •Classical robotics used modular 'sense-plan-act' approaches requiring custom engineering per task with zero knowledge transfer
• •Academia has reached consensus on manipulation methods, with clear signs that scaling data improves performance
• •Previous approaches meant each new task required throwing away all code and starting over

Cheng explains how diffusion policy solved a critical problem in imitation learning - the ability to capture multiple ways of doing the same task while maintaining training stability. This breakthrough enabled scalable data collection from multiple untrained people, rather than requiring the original researcher to collect all data.

• •Diffusion models allow capturing multiple modes of behavior for the same observation while preserving training stability
• •Previous imitation learning was 'finicky' - required the specific researcher to collect data in exactly one way
• •Breakthrough enables multiple untrained people to collect data with different approaches, dramatically improving scalability
• •Unlocked more robust training that doesn't diverge when data shows behavioral variation

Tony describes how ALOHA transformed data collection from unintuitive VR headset teleoperation to a simple, reproducible setup that feels like playing a video game. Combined with ACT (Action Chunking Transformers), this enabled collecting truly dexterous manipulation data and proved transformers work for robotics when you have quality datasets.

• •ALOHA made teleoperation feel like a video game instead of disconnected movement with lag
• •Reduced latency dramatically, enabling faster human reaction and smoother, more dexterous data collection
• •ACT introduced transformers to robotics - they work well with dexterous datasets but failed with shallow networks
• •Action chunking (predicting trajectories vs single actions) mirrors human perception-action cycles and improves consistency
• •Proved transformers architecturally apply to robotics, but only with sufficient data quality

Cheng developed UMI to solve the fundamental bottleneck: teleoperation restricted data collection to labs. By realizing you only need paired observation-action data (video + hand/gripper movement), he created a 3D-printed glove system with GoPros that enabled collecting 1,500 trajectories in two weeks across real-world environments, producing the first end-to-end model that generalizes to unseen environments.

• •Teleoperation setups took hours to configure and restricted data collection to labs, blocking real-world deployment
• •Key insight: you don't need a robot to collect robot data - just video (observation) and hand movement (action)
• •3D-printed glove with GoPro tracks hand position in space and gripper/finger motion through images
• •Three PhD students collected 1,500 clips in two weeks (largest robotics dataset at the time) by taking gloves to restaurants
• •First end-to-end model to generalize to completely unseen environments - robot worked anywhere on Stanford campus
• •Only failure mode: direct sunlight (no sunny data in rainy two-week collection period)

After seeing the UMI robot work across Stanford campus, Tony and Cheng decided to start Sunday Robotics. The company has grown from two founders in an apartment to 30-40 people, bringing together expertise in mechanical engineering, supply chain, software, and controls to build a real consumer product, not just demos. Their mission: put a home robot in every home to eliminate mundane chores and give people time back for family and passions.

• •Company started after UMI demo showed robust generalization - 'maybe we should start a company'
• •Grew from 2 founders to 8 people by end of 2024, now 30-40 people
• •Full-stack approach combines mechanical engineering, supply chain, software engineering, and controls expertise
• •Mission: eliminate mundane chores (dishes, laundry) to give people time for family, hobbies, and passions
• •Focus on building a real product with cost, capability, and robustness - not just research demos

Sunday's design philosophy centers on 'what should a robot look like if it's ubiquitous?' They chose a friendly, cartoon-like face over a Terminator aesthetic. Every design decision prioritizes simplification to accelerate usefulness - like their 3-finger hand that combines three human fingers into one, reducing cost 3x while maintaining most functionality. AI's ability to correct for mechanical imprecision enables using compliant, low-cost actuators instead of expensive, precise industrial components.

• •Design for ubiquity: friendly, cartoon-like face that people see every day, not intimidating Terminator aesthetic
• •3-finger hand design: combines three human fingers into one since they're usually used together (grasping, opening doors)
• •Simplification strategy: multiply cost by 3x for separated fingers vs 1x for combined with most benefits retained
• •AI enables new design space: vision-based correction allows imprecise, compliant, low-cost actuators vs expensive precise industrial robots
• •Industrial robots are fast, stiff, precise because they're blind - AI-enabled robots can be compliant and safe while achieving sufficient accuracy

Sunday plans a 2026 beta program putting various robot prototypes in real homes to learn how people interact with robots and what tasks they prioritize. Current prototypes cost $6k-$20k (mostly cladding at low scale), but will drop under $10k at a few thousand units through injection molding. Commercial availability targeted for 2027-2028, definitely not a decade away, with extremely high standards for safety, capability, and cost.

• •2026 beta program: multiple robot prototypes in selected homes, zero-shot deployment to learn user behavior and preferences
• •Will inform product decisions: do people want to talk to robots? Have robots teach kids? What tasks matter most?
• •Current prototype costs: $6k-$20k, with cladding (few thousand dollars) being the major cost at low scale
• •At few thousand units: cost drops under $10k through injection molding vs CNC and hand painting
• •Commercial availability: 2027-2028 possible, 'definitely not a decade away'
• •Extremely high bar for consumer product: must be highly reliable, capable, and cheap before mass market

Sunday has collected nearly 10 million long-horizon trajectories (not just 'pick up a cup' but full tasks with walking and navigation) using 500+ people with gloves in the wild. The glove system went through 5 major versions with ~20 iterations each (~100 total iterations). Data quality became the critical factor at scale, requiring extensive engineering for automatic calibration, failure detection, and cleaning processes. Glove data surprisingly matches teleoperation quality while enabling more natural, dexterous movements.

• •Nearly 10 million trajectories collected - long-horizon tasks with navigation, not simple pick-and-place demos
• •500+ people using gloves in the wild for data collection, demonstrating massive scale
• •Glove hardware: 5 major versions × 20 iterations each = ~100 total iterations to get production-ready
• •At scale, everything that can go wrong does - people assemble creatively, requiring precise specifications
• •Data quality engineering: automatic calibration before shipping, software to detect hardware failures without human monitoring
• •Glove data quality matches teleoperation after engineering iteration, while encouraging more natural dexterous movement

Sunday sees RL working well for locomotion (easy to simulate rigid body dynamics) but believes imitation learning is faster for manipulation. The key insight: for manipulation, the behavior is easy to learn (close hand with right force) but the world is impossibly hard to simulate (transparent cup with liquid, hand deformation, lighting). For locomotion, it's flipped - world is easy to simulate but reactive behaviors are hard to specify. This makes imitation learning more sample-efficient for their use case.

• •RL excels for locomotion: easy to simulate rigid body contact with ground, but reactive behaviors hard to specify
• •Manipulation is opposite: behavior easy to capture (grasp cup, close hand), world impossibly hard to simulate
• •Simulating transparent cup with orange juice: hand deformation, liquid rippling, rendering, lighting - 'ridiculously hard and expensive'
• •Imitation learning more sample-efficient for manipulation - can learn from demonstrations faster than RL exploration
• •World models theoretically powerful (perfect simulator could cure cancer) but practical question is which method gets there faster
• •Distribution matching critical: robot failed in direct sunlight because two-week data collection was all rainy weather

Two major technical challenges remain: (1) figuring out the training recipe at scale now that sufficient data exists - Sunday is uniquely positioned with their data pipeline, and (2) hardware reliability as the learning team constantly pushes mechanical boundaries. The full-stack approach is essential because standards of 'good' are constantly evolving, making it impossible to work with external partners who can't keep up with changing requirements.

• •Training recipe research: field just entered scaling realm with sufficient data, now need to determine exact recipe for robust behaviors
• •Sunday uniquely positioned due to data volume and complete pipeline for training recipe research
• •Hardware reliability: learning team pushes performance envelope until hardware breaks, then immediate mechanical iteration
• •Hand parts iterated very quickly through tight mechanical-learning team feedback loop under same roof
• •Full-stack necessity: standards of 'good' constantly changing, impossible for external partners to understand evolving requirements
• •Avoid research that doesn't scale: focused on infrastructure and scalable data pipeline before diving into research (started 3 months ago)

Sunday's demos showcase three capabilities: (1) Long-horizon mobile manipulation - cleaning entire messy table, loading dishwasher with precise force control to avoid breaking wine glasses, (2) Zero-shot generalization - tested in 6 Airbnbs with no additional data collection, grasping reflective silverware on transparent tables, and (3) Extreme dexterity - operating espresso machines and folding socks, tasks requiring fine force control impossible with numb teleoperation but natural with gloves where humans feel the forces.

• •Table cleaning demo: mix of fine-grained manipulation with long-horizon full-range mobile manipulation
• •Force control critical: grasping two wine glasses requires precise force (too hard breaks glass, wrong angle shatters in dishwasher)
• •High-stakes tasks: not about recovering from mistakes but not making them in the first place (typical for home tasks)
• •Generalization: 6 Airbnbs, zero-shot, collecting utensils and loading dishwasher with millimeter precision on reflective silverware
• •Transparent tables work because 500+ collectors provided diverse glass table data in training set
• •Dexterous tasks (espresso, sock folding): require fine force control impossible with teleoperation (numb hands can't feel force)
• •Glove advantage: humans naturally feel force closure when fingers touch, enabling intuitive data for force-sensitive tasks
• •Industry firsts: sock folding and end-to-end espresso machine operation

Tony provides a framework for critically evaluating robotics demos: make zero assumptions. Ask if it's autonomous or teleoperated, whether it shows variation (different colored cups, different people), and count the number of interactions (more interactions = more failure opportunities = harder task). Humans naturally project capabilities onto demos, assuming a robot that hands one cup to one person can handle variations, but often it can only do that exact task.

• •Critical questions: Is it autonomous or teleoperated? Does it show variation in objects/people?
• •If they don't show variation (different cup colors, different recipients), robot likely can only do that exact task
• •Humans project capabilities: seeing one cup handoff makes us assume it can do dishes, laundry, etc.
• •Number of interactions matters: each interaction is a failure opportunity, so longer sequences are exponentially harder
• •Only index on what's explicitly shown - don't assume generalization beyond demonstrated scope
• •Visual thinking in robotics creates imagination, but evaluate based on concrete evidence only