Research Intern @ Institute of Automation, Chinese Academy of Sciences (CASIA)
I work on embodied intelligence and world models, robot learning and reinforcement learning, and next-generation agentic architectures for embodied systems.
我目前在中国科学院自动化研究所实习，主要围绕具身智能与世界模型、机器人学习与强化学习，以及面向 Agentic 的下一代具身智能架构开展研究。
- Embodied intelligence & world models — connecting predictive representations with long-horizon planning, causal reasoning, and closed-loop control.
- Robot learning & reinforcement learning — learning robust policies from offline data and continued interaction under real-world constraints.
- Agentic embodied architectures — coordinating perception, memory, planning, tools, and action in open-ended environments.
- FastMCP · merged PR #4704: CodeMode error propagation — Made failed tool calls catchable inside sandboxed CodeMode programs while preserving clear uncaught-error semantics.
- verl · merged PR #7204: per-turn LLM token tracing — Added decoded prompt and response text to asynchronous agent-loop traces for MLflow, Weave, and Trackio.
- AReno · merged PR #473: native attention GPU equivalence — Built numerical checks for varlen prefill and paged decode kernels against an independent PyTorch SDPA reference.
- AReaL · merged PR #1578: TMS launcher reliability — Fixed an
LD_PRELOADconflict that broke multi-GPU TMS offload lifecycle operations. - AReaL · open draft PR #1571: race-safe TMS teardown — Proposed idempotent FSDP cleanup, ordered worker shutdown, and verified process-group termination after offload.
- MERGED · inclusionAI/AReno#473 — test: add native attention GPU equivalence checks
- OPEN · inclusionAI/AReno#472 — docs(agentic): define trajectory observability contract
- OPEN · inclusionAI/AReno#471 — build: adopt file-backed dynamic package version
- OPEN · alibaba/open-code-review#848 — fix(viewer): show running status for active sessions
- MERGED · areal-project/AReaL#1578 — fix(infra): preserve LD_PRELOAD in local launchers
- How can world models become part of the embodied decision loop rather than remain passive prediction modules?
- How should robots combine offline experience with online reinforcement learning to improve safely and continually?
- What abstractions should coordinate perception, memory, planning, tools, and action in the next generation of agentic embodied systems?
Python · PyTorch · FSDP · reinforcement learning · robot learning · RL post-training · distributed systems · MCP