Ep#110: ENPIRE: Agentic Robot PolicySelf-Improvement in the Real World
Wley Atasier, a CMU PhD student, presents EMPIRE with CMU master’s students Jia and Tonha on RoboPapers, explaining how coding agents can improve robot policies through real-world experiments.
In this talk
- 4:03
Why real-world research needs a harness
Atasier distinguishes physical auto-research from approaches that rely on researchers to interpret robot trials and guide agents. The team argues that real-world experiments need safe resets and verifiable rewards so agents can test and improve policies without depending on people as their sensors and actuators. An early attempt to let an agent write an impedance controller broke their robot arm.
- 10:41
Building resets and rewards
Jia describes resets for Push-T and pin insertion, including using OpenCV to locate and reposition the T-shaped object. For zip-tie insertion, Tonha gave the agent segmentation tools, camera views, and success and failure examples to develop a reward function. The resulting classifier reached 80–90% F1 on an offline dataset, and the agent reduced evaluation time from one second to 100 milliseconds.
- 26:21
Agents search under limited budgets
The team ran eight robot workstations, each with an RTX 5090 and about 20 minutes per iteration, using separate coding-agent teams linked through a shared repository. Agents trained a small ResNet-plus-MLP policy after larger models proved too costly to fine-tune within the time limit. The researchers describe the result as budget-limited hill climbing, including a peg-insertion compensation controller that exploited the fixed surface height.
- 34:43
Stronger models and the human role
The team says newer models could make resets and visual learning easier, and describes Astra copying a cup-stacking demonstration, discovering a handover function in the codebase, and adjusting an unstable placement. At scale, Atasier sees people setting goals and requirements while agents handle more implementation, with humans still contributing oversight and verification. The team notes that highly reactive tasks may continue to need low-level controllers.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.