edem-ai/agentchord

AgentChord is the code release for an academic paper called 'From Reaction to Anticipation: Proactive Failure Recovery through Agentic Task Graph for Robotic Manipulation,' accepted at Robotics: Science and Systems 2026. It is a system for controlling robot arms that tries to handle things going wrong while a task is in progress, like an object slipping out of place or falling over. Instead of waiting for the robot to notice failure and react, it tries to predict where the plan can break and bake recovery steps in from the start. The project is built on top of a separate codebase called EmbodiChain. Inside AgentChord, three roles are played by language-model agents. A Task Structuring Agent reads the instruction and what the cameras see, then writes a directed graph of steps for the robot. A Recovery Orchestration Agent looks at that graph, predicts how it might go wrong, defines online checks that detect those failures, and adds branches that move forward into recovery instead of aborting. An Execution Compilation Agent turns the resulting graph into robot motions with monitors that can interrupt motion when something looks wrong. AgentChord can run in a GPU simulation or on a real robot called CobotMagic. The README focuses on the simulation. It needs Ubuntu 20.04 or newer with an NVIDIA GPU at compute capability 7.0 or higher, and Python 3.10 or 3.11. Install is a git clone plus pip install with an extra package index, and a Docker image based on Ubuntu 22.04 and CUDA 12.8 ships the full simulation stack. The agents call any OpenAI-compatible endpoint via LangChain, with gpt-5 as the default model name. A simulation rollout is started with a python command that points to gym and agent config files and a task name like SinglePourWater. While it runs you can press f in the terminal to inject a failure such as a misplaced or fallen object, and the system tries to recover. The generated task graph, recovery spec, and compiled graph are cached as JSON files under embodichain/database/agent_generated_content/sim/, and the cache is reused unless you pass --regenerate. The repository has 14 stars and is written in Python.