AstrAlignment
AstrAlignment is a human-agent-robot evaluation dashboard that tests whether an agent’s physical-world assurances are supported by evidence.
Core trick
Forks simulated robot world state at human-handoff points to test whether an agent's physical-world claims hold, then exports the failure episodes as ROS bags.
Description
AstrAlignment is a human-agent-robot evaluation dashboard that tests whether an agent’s physical-world assurances are supported by evidence. It simulates robot handoffs and forks the world state to evaluate what happens when a human follows, delays, or cancels an instruction. Astra coordinates tasks through structured actions, analyzes execution traces, and generates bounded counterfactual scenarios that expose policy failures. Researchers can replay failures, compare policy versions, inspect supporting evidence, and export reproducible episodes as ROS bags for policy evaluation and training. Future work will use these verified failure-and-recovery traces to support continuous policy improvement and self-improving robotic systems.
README excerpt
# Astralignment > This README documents the existing prototype and its historical research direction. > The product name is undecided. The [first-iteration specification](docs/OSS_SPEC.md) > defines evidence curation for robot-policy training teams, with a CPU-only workflow > and no required model API key. A [local curation preview](docs/CURATION_IMPLEMENTATION.md) > implements the local end-to-end workflow at /curation; it has not passed all release > gates. It does not change the prototype setup or model requirements below. **A human–agent–robot evaluation workbench for testing whether an agent's assurance is supported by the physical state.** An agent's statement can change what a person does next. If the agent says a robot has secured an object, a person may remove the support beneath it. The relevant question is not whether the explanation sounds reasonable, but whether the object remains supported after the person acts. Astralignment is being developed for robotics, embodied-AI, and alignment researchers evaluating policies under human reliance. The target is not dexterity. It is the coordination boundary between perception, policy execution, verification, and human action. The current implementation provides two simulated Unitree G1 humanoids, MuJoCo checkpoint forks, independent coordination checks, an Astra repair loop, real-video overlays, and durable JSON/MCAP experiment records. It tests crossing and right-of-way failures. No physical robot is controlled. **O