Self-Evolving Coding Agents for the Physical World: HexaAnything Bridges Code as Policy to Autonomous Robotics

核心背景与行业痛点

Vision-Language-Action (VLA) and World-Action Models (WAM) govern modern robotics by directly mapping camera frames and text goals to joint velocities or end-effector trajectories. However, implicitly encoding world dynamics, task progress, and exception handling inside low-level continuous action vectors renders policies brittle: slight alterations in object poses or camera viewpoints trigger catastrophic execution crashes. Crucially, raw action streams prevent high-level symbolic reflection, debugging, or reusable procedural recall.

架构亮点与底层机制

Researchers introduce Physical Coding (arXiv:2609.35432), porting digital software agent principles into embodied physical workflows:

1. Code as World: Represents spatial geometries, object relations, and dynamic progress metrics explicitly as structured, queryable programmatic data structures.
2. Code as Policy: Formulates planning, execution loops, safety verifications, and exception rollbacks as executable Python routines that orchestrate specialized tools alongside low-level VLA control primitives.
3. HexaAnything Autonomous Orchestrator: Integrates multimodal perception, LLM reasoning, and robotic execution. In-the-loop sensory feedback enables real-time code rewriting, while verified successes crystallize into reusable procedural memory.
4. Hierarchical Self-Evolution: Synthesizes execution traces into curated post-training datasets to update model weights (HexaModel), optimize tool harnesses, and iteratively refine physical task protocols.

权威 Benchmark 与实测跑分对比

Benchmarked across simulated environments, physical reasoning suites, and physical dual-arm robotic systems:

1. Outperforms XR-1 VLA on RoboCasa365: Establishes clear superiority across Composite-Unseen scenarios and overall success metrics on RoboCasa365.
2. Weight Internalization Superiority: The harness-trained HexaModel systematically outperforms foundation baselines across every task split.
3. Autonomous Physics Experimentation on Dual-Arm AgileX Robots: Deployed on physical hardware, HexaAnything autonomously executes complex physics experiments on PhyBench and diverse tabletop manipulation workflows with lower latency than published baselines.

开发者实战落地与开箱指南

The HexaAnything codebase, simulation harnesses, and robotic control bridges are publicly released. Robotics teams can deploy Physical Coding as an interpretable symbolic middleware over low-level motor primitives, transforming robotic manipulators into self-correcting autonomous coding agents capable of continuous physical self-evolution.