New agentic embodied‑AI papers

A robot that can pick up a mug is old news in research labs. A robot that can notice the mug is upside down, revise its plan, ask a different module for help, and recover after a bad move is what this new wave of papers is trying to build. (arxiv.org) Embodied artificial intelligence means software that does not stop at answering questions on a screen. It sees through cameras, acts through motors, and pays a price for mistakes because the world pushes back. (nature.com) That is why robots have lagged behind chatbots. A language model can bluff its way through a paragraph, but a robot that misreads one object or misses one step can fail an entire kitchen task. (arxiv.org) Older robot systems often split the job in two. One part made a plan in words, and another part handled motion, which worked for short demos but broke down when a task stretched across many steps. (arxiv.org) The new idea is agentic orchestration. Think of it like a movie director calling on separate crews for vision, memory, planning, action, and checking, instead of asking one exhausted actor to do every job at once. (arxiv.org) One of the newest examples is a paper called RoboAgent, submitted on April 10, 2026. It says a single vision-language model should actively invoke different sub-capabilities, each with its own context, while a scheduler decides what gets called next. (arxiv.org) In plain English, RoboAgent tries to turn a fuzzy chain-of-thought robot into a checklist robot. The paper breaks a long task into smaller vision-language problems so the system can produce intermediate results before it commits to the next action. (arxiv.org) A 2025 paper called Agentic Robot pushes the same direction from the manipulation side. It uses what the authors call Standardized Action Procedures, modeled on the written procedures human teams use in factories and hospitals. (arxiv.org) That framework assigns three jobs to three parts: a reasoning model breaks an instruction into subgoals, a vision-language-action model turns camera input into control commands, and a temporal verifier checks whether the robot should continue or recover. On the LIBERO benchmark, the paper reports a 79.6% average success rate, beating SpatialVLA by 6.1 points and OpenVLA by 7.4 points on long-horizon tasks. (arxiv.org) Another strand of the field is giving robots memory outside the model itself. A Computer Vision and Pattern Recognition 2024 paper on Retrieval-Augmented Embodied Agents adds a policy memory bank so a robot can look up similar past situations instead of relearning from scratch every time. (openaccess.thecvf.com) A separate March 16, 2026 Nature Machine Intelligence paper moves this closer to deployment. It connects a large language model agent to the Robot Operating System, translates model outputs into robot actions, supports behavior trees and inline code, and releases the framework as open source. (nature.com) The newest twist is that some researchers now treat oversight as its own runtime layer, not a side note inside the agent. A paper posted on April 9, 2026 reports a governance layer that checks policy, monitors execution, handles rollback, and allows human override, with 96.2% interception of unauthorized actions and 91.4% recovery success in 1,000 randomized simulation trials. (arxiv.org) Put together, these papers are less about a single breakthrough robot and more about a new blueprint. The field is moving from “can a model output actions” to “can a system keep perceiving, planning, executing, checking, and recovering until the whole job is done.” (arxiv.org)