Stanford Robotics Seminar ENGR319 | Winter 2026 | Gen Control, Action Chunking, Moravec’s Paradox

Why It Matters

Understanding and overcoming the algorithmic Moravec's paradox is essential for scaling reliable robot manipulation, directly impacting industrial automation and the commercial viability of data‑driven control systems.

Key Takeaways

• Moravec's paradox now has algorithmic counterpart limiting robot learning.
• Action chunking and generative control policies enable recent robotics breakthroughs.
• Behavior cloning suffers exponential error growth due to horizon and stability issues.
• Offline imitation learning cannot overcome compounding errors without richer policy representations.
• Reparameterizing closed-loop dynamics is crucial for scaling data-driven robot control.

Summary

The Stanford Robotics Seminar examined why learning from demonstration remains harder for physical robots than for symbolic AI, coining an "algorithmic Moravec's paradox" that highlights fundamental instability in continuous control. The speaker traced the recent surge in narrow manipulation capabilities to two algorithmic breakthroughs—action chunking, which predicts sequences of actions for open‑loop execution, and generative control policies that model multimodal action distributions.

A central argument was that behavior cloning, despite its supervised‑learning appeal, incurs exponential compounding error because the learned policy’s closed‑loop dynamics can become unstable, a problem formalized through a negative result showing any smooth, Markovian learner will suffer horizon‑dependent degradation. The talk illustrated these concepts with a student’s bimanual tele‑operation folding demo and highlighted the 2023 inflection point when data scaling combined with these algorithmic tricks sparked industrial interest.

Key examples included the mathematical proof that even with perfect square‑loss fitting, error scales as n⁻ᵅ yet rewards diverge exponentially, and the comparison to discrete settings where errors only grow polynomially. The speaker emphasized that reparameterizing the robot‑learner closed‑loop system—essentially reshaping how policies interact with dynamics—unlocks the "bitter lesson" of scaling data‑driven control.

The implications are clear: future robotics progress hinges on designing policies that anticipate and mitigate instability, rather than merely collecting more data. Companies seeking robust autonomous manipulation must adopt action‑chunking and generative control frameworks, and researchers must explore richer, possibly non‑Markovian representations to break the compounding‑error barrier.

Original Description

For more information about Stanford’s graduate programs, visit: https://online.stanford.edu/graduate-education

February 27, 2026

This seminar covers:

• Mathematical evidence for Moravec’s claim

• Key innovations in modern robot learning

To follow along with the seminar schedule, visit: https://stanfordasl.github.io/robotics_seminar/

Max Simchowitz is an Assistant Professor at the Machine Learning Department at CMU.