Zhicheng Zheng

We ask whether coding agents can solve long-horizon, dexterous robotics tasks and turn those solutions into scalable supervision for robot policies.

EmbodiedSWE-Bench spans contact-rich manipulation, deformable objects, and tasks needing up to half an hour of continuous interaction; EmbodiedSWE-Gen expands a single agent solution into large, diverse trajectory sets for training a VLA—which, trained only on coding-agent-generated simulation data, completes a long-horizon task on a real robot.

An agentic system for robotic question-answering and navigation that stores visual embeddings with pose and time in a vector database, grounding retrieval in a spatial map to answer queries with navigation goals.

It retains over 97% of its performance on frontier base models and still preserves 67% on small open-source models, while achieving over 250× storage compression without accuracy loss and 10× greater retrieval efficiency compared to a naïve VLM.

We introduce a novel benchmark for assessing machine physical commonsense on various continuum bodies, encapsulating conceptual inference and dynamic reasoning.

Contributes to:
ByteDance Seed 2.0
PhysBench