Nov 19, 2025 · 39m · no-priors
No Priors Ep. 141 | With Sunday Robotics Co-Founders Tony Zhao and Cheng Chi
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Sunday Robotics co-founders Tony Zhao and Cheng Chi join No Priors to discuss developing Memo, a safe and affordable domestic robot designed to automate household chores through full-stack hardware engineering and scaled imitation learning.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 14.2% of the talking time here. How this is scored →
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Tony dismisses the popular thesis that universal world simulators will solve manipulation, calling it a tautology and explaining why simulation fails when modeling nuanced real-world fluid and deformation dynamics.
Hardest push from the hosts ▶ 13:03 Sarah Demands Justification for Non-Humanoid ArmSarah directly questions Sunday's engineering choice, asking why the obvious answer isn't simply replicating a full five-finger humanoid arm.
Biggest teaching moment ▶ 22:35 Inverted Complexity: Locomotion vs. ManipulationCheng educates the audience on why RL works in locomotion (simple flat point contact physics) while manipulation has flipped difficulty (simulating orange juice in a glass is nearly impossible, but human imitation learning is straightforward).
The host holds their own ▶ 6:06 Connecting Action Chunking to Transformer Token SequencesSarah demonstrates expert domain mastery by connecting robotic action chunking trajectories directly to autoregressive sequence prediction in large language models.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The State of AI Robotics: Between GPT and ChatGPT | 4 | 4 | 0 | 0 | Sarah sets the context by asking why classical robotics struggled for decades and why foundation models offer renewed optimism. The guests explain the gap between GPT (technology) and ChatGPT (product) and critique the non-generalizable sense-plan-act paradigm. | |
| Breakthroughs in Imitation Learning via Diffusion Policy | 6 | 4 | 0 | 0 | Sarah demonstrates strong technical domain knowledge, asking specifically about Diffusion Policy, Aloha, ACT, and the analogy between action chunking and LLM token sequence prediction. The guests explain how diffusion models solved multimodal imitation learning. | |
| The UMI Gripper, In-the-Wild Data, and Founding Sunday | 5 | 4 | 0 | 0 | Cheng recounts inventing the 3D-printed UMI gripper to bypass bulky lab teleoperation setups using GoPros in the wild. Sarah highlights the remarkable capital efficiency of their academic breakthroughs compared to current scaling efforts. | |
| Sunday's Mission and Full-Stack Team Growth | 5 | 4 | 1 | 4 | Sarah pushes back on Sunday's simplified hardware architecture, asking why they did not opt for a standard five-finger human arm. The guests justify their three-finger design and compliant low-cost actuators enabled by vision-based AI corrections. | |
| Product Roadmap, Generalization Demos, and Beta Testing | 4 | 3 | 0 | 1 | The guests outline their 2026 home beta testing roadmap and discuss the cross-product of dexterity and generalization. They note that scaling data yielded surprising zero-shot precision on metallic forks and ceramic plates. | |
| Full-Stack Iteration and Scaling to 10 Million Trajectories | 4 | 3 | 0 | 0 | Sarah asks about training data scale relative to the rest of the robotics industry. The guests explain their 100+ hardware iterations and reveal they have collected nearly 10 million long-horizon wild trajectories with over 500 data collectors. | |
| Evaluating Scaling Paradigms: Imitation Learning vs. Simulation and RL | 5 | 6 | 1 | 2 | Sarah probes alternative scaling paradigms including RL and world models. The guests give a deep pedagogical contrast showing why simulation works for locomotion but fails for manipulation due to the intractable complexity of real-world rendering and fluid dynamics. | |
| Managing Data Quality and Automated Failure Detection | 4 | 4 | 0 | 0 | Sarah asks what views have changed over the past year. The founders emphasize automated hardware calibration, failure detection systems, and intentionally avoiding academic research tricks until scalable data infrastructure was fully operational. | |
| Commercial Deployment Timelines and Consumer Pricing Targets | 4 | 3 | 0 | 3 | Sarah presses the founders for uncomfortable estimates on consumer availability and pricing. The guests detail a 2027-2028 commercial target and explain how shifting from CNC cladding to injection molding brings BOM costs under $10,000. | |
| The Societal Impact of Zero-Cost Domestic Labor | 5 | 5 | 1 | 1 | Sarah and the guests discuss the societal impact of zero marginal cost domestic labor. The founders share a critical rubric for evaluating viral robotics demos, warning viewers to make zero assumptions about autonomy and distribution generalization. | |
| Sunday's Benchmark Demos: Table Clearing, Airbnb Testing, and Sock Folding | 4 | 5 | 0 | 0 | The guests describe Sunday's core launch benchmarks: full table cleanup with fragile wine glasses, zero-shot Airbnb testing on reflective cutlery, and delicate tactile tasks like sock folding requiring fine force closure. | |
| Hiring Full-Stack Roboticists and Sunday's Culture | 3 | 2 | 0 | 0 | The conversation concludes with hiring priorities for full-stack roboticists. Cheng shares his personal progression from mechanical engineering to programming and machine learning. |
Statements from this episode (0)
Nothing in this episode matches those filters. clear them