No PriorsNo Priors Ep. 141 | With Sunday Robotics Co-Founders Tony Zhao and Cheng Chi
CHAPTERS
- 0:00 – 0:55
Why home robots now: cheap, safe, capable—and a billion-unit vision
Tony frames the core consumer value proposition: eliminate household chores so people can spend time on what they care about. The founders set an ambitious north star—ubiquitous home robots—grounded in cost, safety, and capability.
- •Household chores (dishes, laundry) as the killer use case
- •Three product constraints: cheap, safe, capable
- •Belief in mass adoption: potentially 1B+ robots in homes
- •Robot as a consumer product, not just a lab demo
- 0:55 – 2:11
Robotics is between “GPT” and “ChatGPT”: recipe emerging, product not yet scaled
They contextualize current AI robotics as having early signs of a scalable learning recipe, but not yet the scale or polish that yields a mainstream consumer experience. The big open question is what happens when robotics data and training scale up dramatically.
- •Analogy: GPT (technology) vs ChatGPT (product)
- •Academic consensus forming around methods for manipulation
- •Performance expected to improve with data/model scaling
- •Robotics lacks the equivalent of “internet-scale” data
- 2:11 – 3:13
Why classical robotics progressed slowly: brittle sense-plan-act and task-by-task engineering
Cheng explains how the traditional modular stack required bespoke interfaces per task and environment, creating little reuse. Both academia and industry often rebuilt systems from scratch for each new application, leading to slow, non-compounding progress.
- •Sense-plan-act modular pipelines with human-designed interfaces
- •One-task/one-environment engineering that doesn’t generalize
- •Academic incentives: each task becomes a separate paper
- •Industrial systems optimized for narrow applications
- 3:13 – 4:29
Diffusion Policy’s impact: making imitation learning stable across diverse demonstrations
They introduce diffusion policy as a major unlock for imitation learning by handling multimodal behaviors without destabilizing training. This reduces the need for a single expert collector and makes data collection more scalable and tolerant to variation.
- •Imitation learning: supervised learning on observation–action pairs
- •Historical pain: finicky training, requires “one way” to do a task
- •Diffusion models capture multiple valid modes of behavior
- •Enables data collection from multiple (even untrained) people
- 4:29 – 7:01
ACT + ALOHA: intuitive teleop, better dexterous data, and transformers that finally work
Tony explains how ALOHA made teleoperation feel more natural (less delay), improving dexterous data quality and enabling harder tasks. With stronger datasets, transformers and system techniques like action chunking became viable for manipulation.
- •ALOHA improves teleop usability and reduces lag
- •Dexterous data enables investigation of harder manipulation problems
- •Transformers become effective once data is strong enough
- •Action chunking: predict trajectories vs millisecond-level actions
- •Chunking yields smoother, more human-like, consistent behavior
- 7:01 – 10:38
UMI: collecting robot-like data without robots using a GoPro + tracked gripper
Cheng describes UMI as a breakthrough in data collection: capture observation and “action” (hand/gripper motion) from video and tracking, outside the lab. This enabled rapid dataset growth and generalization to unseen environments (with distribution caveats like sunlight).
- •Teleop setups are slow to deploy and lab-bound
- •Key insight: robotics data = paired observation (video) + action (hand motion)
- •GoPro-based tracking can approximate action signals
- •3 people collected ~1,500 clips quickly; became a large robotics dataset
- •Generalization demo success—and failure under unseen sunlight highlights distribution matching
- 10:38 – 11:57
From Stanford cart demos to Sunday: building a full-stack robotics company
A strong campus generalization moment convinced them to start Sunday, evolving from a two-person apartment setup to a growing team. They emphasize that productizing robotics requires integrating mechanical engineering, controls, software, and operations—not just publishing demos.
- •Company origin story: prototypes in an apartment, rapid team growth
- •Full-stack integration: mech, supply chain, controls, software, ML
- •Goal: real product reliability vs one-off lab demos
- •Robotics scaling requires tight iteration across teams
- 11:57 – 14:03
Design philosophy for a ubiquitous home robot: simplify for cost, safety, and repairability
They discuss how consumer ubiquity changes industrial design choices: friendliness (a “face”), pragmatic form factors, and simplification over full humanoids. Examples include a three-finger hand with coupled fingers to cut cost while preserving utility.
- •Ubiquitous robots should feel friendly and approachable
- •Humanoid isn’t automatically optimal if it slows shipping
- •Simplify wherever possible to accelerate time-to-useful product
- •Coupled three-finger design: most use is collective, avoids 3× complexity
- •Design for low cost and easy repair
- 14:03 – 15:05
AI changes hardware requirements: vision-enabled robots can use cheaper, compliant actuators
Cheng contrasts blind industrial robots (stiff, precise) with AI-perception robots that can correct errors in real time. This opens a design space for compliant, safer, lower-cost hardware that still reaches sufficient home-task accuracy via software.
- •Industrial robots are blind → require stiffness and precision
- •Perception enables closed-loop correction of mechanical inaccuracies
- •Compliant actuators: cheaper and inherently safer, but imprecise
- •System + learning compensates for hardware imperfections
- •Home tasks need “sufficient” accuracy, not industrial overkill
- 15:05 – 16:42
Shipping plan and milestones: 2026 beta, consumer launch dependent on reliability
Tony outlines plans to run beta programs in 2026 to learn real in-home usage and refine product requirements. They stress high standards—safety, capability, and low cost—and indicate a potential commercial timeline of 2027–2028 depending on beta results.
- •Many internal prototypes already exist
- •2026: in-home beta programs to observe real user behavior
- •Learning goals: interaction patterns, desired chores, product shape
- •Hard constraints: extremely safe, capable, low cost
- •Commercial timing likely 2027–2028; not “a decade away”
- 16:42 – 19:34
Scaling data in the wild: the glove pipeline, 10M trajectories, and why quality becomes everything
They describe the operational reality of scaling: iterative glove versions, calibration, automatic failure detection, and data cleaning. Sunday reports collecting nearly 10 million long-horizon trajectories in the wild, and they highlight how data quality demands robust processes at scale.
- •Scaling is “long and painful”: reliability and QA across hardware + pipeline
- •Glove iteration: V0→V5 with many sub-iterations; real-world usage reveals edge cases
- •500+ people collecting data in the wild; assembly and failures must be controlled
- •~10M trajectories including navigation + long-horizon tasks, not single-action snippets
- •Data quality requires calibration, automated checks, filtering, and cleaning
- 19:34 – 24:55
How to scale learning: glove data vs teleop, RL’s role, and why simulation is harder for manipulation
They compare scaling approaches—teleop, glove data, RL, and world models—focusing on which gets to product faster. They argue RL shines for local motion where simulation is easier, while imitation learning excels in manipulation because behavior is easier to learn than perfectly simulating complex objects and rendering.
- •Initial worry: glove data higher quantity but lower quality than teleop
- •Finding: glove form factor yields more natural dexterous motions; quality can match after engineering
- •Key challenge: convert human-in-frame data to robot-equivalent signals
- •RL: powerful but less sample-efficient; best for local motion with tractable simulation
- •Manipulation: hard to simulate the world (contacts, deformables, liquids, transparency), easier to learn behavior from demos
- 24:55 – 30:47
Remaining technical challenges: scalable training recipes and hardware reliability loops
They identify two main frontiers: discovering robust training “recipes” that work at scale, and pushing hardware reliability while learning keeps stressing the system. Full-stack co-location accelerates the break–fix–redesign loop between learning and mechanical teams.
- •Field is newly entering true scaling regimes; recipes aren’t settled
- •Need research that scales, not “cute” ideas that won’t transfer to product
- •Hardware is still being pushed to its limits; failure is part of iteration
- •Co-located teams enable rapid iteration from breakage to redesign
- •Full-stack approach reduces dependency on external partners’ shifting standards
- 30:47 – 36:53
Why robotics demos mislead—and what Sunday’s demos aim to prove (autonomy, long horizon, generalization)
Tony explains how to evaluate demos skeptically: check autonomy vs teleop and whether variations/generalization are shown. They then describe Sunday’s demo suite: long-horizon table cleanup and dishwasher loading, zero-shot Airbnb generalization, and fine force-control tasks like espresso and sock folding.
- •Demo checklist: autonomy vs teleop, and shown scope vs assumed scope
- •Long sequences matter: more interactions increase failure probability
- •Demo category 1: full messy-table cleanup + dishwasher operation (mobile manipulation + long horizon)
- •Demo category 2: zero-shot generalization in new homes (Airbnbs), precision utensil/plate tasks
- •Demo category 3: dexterity/force tasks (espresso machine, sock folding) enabled by natural force-aware data collection
- 36:53 – 39:16
Hiring and culture: training full-stack roboticists to build the complete system
They close with what they’re hiring for: people who can span disciplines or want to grow into full-stack roboticists. The founders emphasize robotics as an end-to-end product discipline—hardware, software, data, and user experience must be designed together.
- •Hiring focus: full-stack roboticists (or aspiring full-stack)
- •Robotics requires mechanical, electrical, controls, software, and data skills
- •Company teaches/mentors software engineers into robotics roles
- •Work spans consumer product + ML + hardware integration
- •Emphasis on building the whole experience, not siloed components
