Stanford Robotics Seminar ENGR319 | Spring 2026 | Ingredientsfor Long-Horizon Robot Autonomy

Why It Matters

Enabling long‑horizon autonomy with efficient memory and reliable primitives will expand robot utility across home and industrial sectors, driving broader adoption and new revenue streams.

Key Takeaways

• Robots excel at short, dexterous tasks but lack long‑horizon autonomy.
• Memory and high‑success primitives are essential for extended task execution.
• Adding raw visual history slows inference and introduces distribution shift.
• Multiscale memory compresses images short‑term, uses language long‑term.
• New papers MEM and PIO7 aim to boost generalization and robustness.

Summary

The Stanford Robotics Seminar highlighted Physical Intelligence’s push toward truly autonomous, long‑horizon robots that can handle everyday home and industrial jobs. While recent advances enable robots to perform complex, short‑duration tasks—like unlocking a lock or precise object reorientation—the speaker emphasized that these are isolated steps, not full‑scale jobs that humans assign.

Key challenges identified include the lack of persistent memory, insufficiently reliable primitive skills, and poor generalization across varied environments. Without memory, robots repeat actions endlessly, as illustrated by a plate‑washing loop or a grocery‑bag unpacking scenario where the robot forgets whether items remain inside. Moreover, low‑success rates for individual skills compound over extended sequences, making long‑duration autonomy impractical.

The talk introduced two recent papers: MEM (Multiscale Embodied Memory) and PIO7. MEM proposes a dual‑memory architecture—dense, compressed visual snapshots for short‑term context and abstracted language representations for long‑term tracking—addressing latency and distribution‑shift issues that arise when naïvely feeding raw histories into policies. PIO7 focuses on improving skill robustness and generalization, ensuring that each primitive operation meets high success thresholds before being chained together.

If successful, these innovations could transform robotics from task‑specific tools into versatile assistants capable of cleaning apartments, assembling server racks, or managing supply chains autonomously, unlocking new market opportunities and reducing the need for costly retraining whenever environments change.

Original Description

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

April 17, 2026

This seminar covers:

• Developing the ingredientsfor long-horizon robot autonomy, which are giving robot policies a sense of memory, and training generalist behaviors that are both broadly capable and high-performing

• Recent progress on both fronts through two works: π0.6-MEM and π0.7

• What's still missing on the path to long-horizon physical autonomy

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

Karl Pertsch is a member of technical staff at Physical Intelligence, where he works on training robot foundation models.