Agents that Act, Build, and Evolve.
A curated reading list for robot execution, robotics development, and persistent improvement.
Foundation-model agents can direct robot behavior, build robotics workflows, and revise systems from feedback. This collection follows the framework of our forthcoming paper, Agents that Act, Build, and Evolve: A Unified Framework and Platform for Agentic Robotics.
The three reading lists group works by the decisions made by the agent. A work can appear more than once when its configurations differ.
| Role | What the agent does | Browse |
|---|---|---|
| Act — Agent as Policy | Select behaviors, orchestrate tools, or write robot-execution programs. | Observation × execution interfaces |
| Build — Agent as Infrastructure | Construct and coordinate collection, training, evaluation, and diagnosis workflows. | Development lifecycle |
| Evolve — Agent as Self-Evolver | Use feedback to revise and retain artifacts, components, or the agent harness. | Retained update targets |
| Observation interface | A1: Behavior Selection | A2: Tool Orchestration | A3: Code Synthesis |
|---|---|---|---|
| O1 — Sensory input | Code executor · Model executor | O1 / A2 | O1 / A3 |
| O2 — Perception tools | Code executor · Model executor | O2 / A2 | O2 / A3 |
| O3 — Privileged state | Code executor · Model executor | O3 / A2 | O3 / A3 |
| Stage | Categories |
|---|---|
| Collection | Environment Generation · Dataset Construction |
| Training | Training |
| Inference | Evaluation · Diagnosis |
| Family | Update targets |
|---|---|
| Iterative Artifact Refinement | Programs and Skills · Knowledge and Memory · Rewards · Environments and Datasets · Robot Morphology |
| Component Optimization | Model · Tool Code |
| Harness Improves Harness | Representation · State Management · Tool Interface · Execution Workflows |
| Successor AI R&D | Research agents and succession |
We include documented foundation-model agent decisions in physical or simulated robot tasks and robotics development. Ordinary VLA inference, implicit reasoning, fixed scoring, and fixed training alone do not establish an agent role. Classification follows the demonstrated configuration.
Each entry links to a primary paper, report, project, or repository. Source types are labeled. A clickable 
in the Code column links to available implementation code; a dash means no code link is currently listed. A literature entry does not imply that its implementation has been integrated into RoboAgentLab or that its reported results have been reproduced.
- RoboAgentLab platform — Robotics extensions for the pi agent harness. Services expose simulation, robot, perception, planning, and policy capabilities; skills describe how to use them.
- Robot policy interactions — Agent interactions with robot environments.
- Robotics R&D interactions — Coding-agent traces from robotics development and testing.
- Paper — Public manuscript and citation information will be linked here when released.
Corrections and additions are welcome. Please provide a primary source, the relevant robot-domain agent decision, and the proposed classification. See CONTRIBUTING.md.
The searchable metadata is in data/papers.json; reading lists can be regenerated with python3 scripts/build.py. This collection was initialized from our paper's taxonomy on 20 October 2026.
We thank Awesome-Agentic-Robots for inspiration for the reading-list presentation. This collection uses RoboAgentLab's role-based taxonomy.
Original list text and metadata are available under CC BY 4.0. Linked papers, code, logos, and datasets retain their respective licenses.