Agentic Robotics
Linzhan Mou ·
What interests me most about robot agents is their ability to learn from failures. If a cup slips during a grasp, an agent can inspect the video and execution logs, form a hypothesis about what went wrong, and test a change. Recent work gives agents tools to carry out this process and retain useful fixes as code, skills or memory.1, 3, 10, 12, 13
I am less convinced that practicing a skill in simulation is enough to make it reliable on a real robot. A simulated grasp may work because the friction is too high, the object is too light, or the contact geometry is wrong. More practice could make the policy better at exploiting those assumptions. Even a scene that looks right can behave differently when the robot touches it.
This makes real-world failures and demonstrations especially useful. A failed attempt shows where the current system breaks; a demonstration provides an example of an interaction that works. Real-to-sim reconstruction could use both to build practice environments around those interactions, then compare simulated rollouts with the recorded behavior.2, 4, 5
I would like to see agents use this feedback to improve the simulator and policy together. After a failure, the agent could test whether changing the policy helps, whether the simulator needs correction, or whether both do. SimEX takes a step in this direction by using physical trials to update the simulator and robot toolbox.6 Improvements might become reusable code, recovery routines, or data for training a learned policy.8, 9
The next real-world trial is the important check. Does the change prevent the failure under new conditions, with fewer retries and less human help? Repeatable resets and independent evaluation make that question easier to answer.7, 11 That is the kind of progress I would want an agent to keep.
Blogs
01 /
Gemini 4 Argon for Robotics
A general-purpose physical reasoning model from Google, shown writing robot control code, building MuJoCo scenes from robot views, human video and sketches, labeling data quality and progress, and improving actions through simulated practice. It scores 87% on the ERQA embodied reasoning benchmark.
02 /
The used key is always bright.
Haozhe Jiang on why agentic robotics needs visual perception and efficient iteration, not just cheap code generation, and how RPG turns offline demonstrations into a practiced skill library that keeps improving while the foundation model stays frozen.
Cites: RPG, Playful Agentic Robot Learning, ASPIRE, CaP-X.
Papers
01 / Added
RAX: Scaling Online Exploration for Robot Agents
A frozen coding agent explores directly on real and simulated robots without demonstrations, teleoperation or policy training, then compiles what works into verified programs and skills kept in its workspace. Generalization improves as exploration scales across single-arm, bimanual, dexterous, mobile and humanoid robots.
02 /
Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
Project · Code coming soon · Paper
Uses offline demonstrations to reconstruct simulation practice tasks, then diagnoses failures and refines skills and system prompts. Cross-task validation retains reusable improvements before the frozen execution system is deployed on real robots.
Cites: ASPIRE, Playful Agentic Robot Learning, CaP-X, ENPIRE.
03 /
Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation
Project · Code not released · Paper
Rebuilds real scenes as digital twins where an agent explores failures and writes recovery skills, run directly as code or used to train recovery policies. Real rollouts invoke those skills or ask for human help, and DAgger rounds turn the recovered episodes into data that improves the base VLA policy.
Cites: CaP-X, ASPIRE, ENPIRE, Playful Agentic Robot Learning.
04 /
SimEX: Simulation-Integrated Robotics AutoResearch
Project · Code not released · Paper
Uses simulation as an experimental laboratory where coding agents probe and improve a robot toolbox. A few physical trials then update both the simulator and toolbox, allowing simulated diagnosis and repair to support real-world deployment.
Cites: CaP-X, ASPIRE, Playful Agentic Robot Learning, Harness VLA, ENPIRE.
05 /
RLE-Bench: A Qualifying Exam for Coding Agents as Robot Learning Engineers
Evaluates coding agents as robot learning engineers across interactive control, policy learning, perception and estimation, and mechanical design. Its contribution is measuring broader engineering capability through independently evaluated artifacts under resource budgets.
06 /
EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics
Benchmarks coding agents on long-horizon, dexterous manipulation. EmbodiedSWE-Gen turns a verified program into diverse demonstrations for VLA training, connecting task-specific code solutions to learned robot policies.
07 /
Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents
Uses a frozen VLA as a retryable contact-rich primitive, coordinated by a planner with memory and analytic controls. Its contribution is extending VLA capabilities through staging, re-grounding, and recovery without fine-tuning the VLA.
Cites: CaP-X, ASPIRE, Playful Agentic Robot Learning.
08 /
ASPIRE: Agentic /Skills Discovery for Robotics
Writes and repairs robot-control code using detailed execution traces, then stores validated fixes as reusable skills. Its core contribution is a continual skill-discovery loop combining feedback, an expanding skill library, and evolutionary program search.
Cites: CaP-X.
09 /
ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
Builds a repeatable real-world loop of automatic reset, policy execution, outcome verification, and improvement. Coding agents use physical rollout feedback to revise training infrastructure and algorithm code, enabling autonomous policy improvement on individual robots or a robot fleet.
Cites: CaP-X.
10 /
Playful Agentic Robot Learning
RATs proposes its own practice tasks, writes and tests robot programs, and distills successful behavior into reusable code skills. Self-directed play builds a skill library before downstream tasks arrive, without updating model weights.
Cites: CaP-X.
11 /
CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation
Introduces CaP-Gym and CaP-Bench to study coding agents for robot manipulation. CaP-Agent0 improves execution through agentic feedback without training, while CaP-RL uses verifiable rewards to improve policies and transfer them to real robots.
- Rao, K. (2026). Gemini 4 Argon for Robotics. Google Developers Blog.
- Jiang, H. (2026). The used key is always bright.
- Huang, R. et al. (2026). RAX: Scaling Online Exploration for Robot Agents.
- Wang, Y.-J. et al. (2026). Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents.
- Liu, I. et al. (2026). Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation.
- Hu, J. et al. (2026). SimEX: Simulation-Integrated Robotics AutoResearch.
- Ma, H. et al. (2026). RLE-Bench: A Qualifying Exam for Coding Agents as Robot Learning Engineers.
- Shen, Z. et al. (2026). EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics.
- Zhang, Y. et al. (2026). Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents.
- Lu, R. et al. (2026). ASPIRE: Agentic /Skills Discovery for Robotics.
- Xiao, W. et al. (2026). ENPIRE: Agentic Robot Policy Self-Improvement in the Real World.
- Zhang, J. et al. (2026). Playful Agentic Robot Learning.
- Fu, L. et al. (2026). CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation.