I am Yibo Li (李一博), a master’s student in Software Engineering at Beihang University, advised by Prof. Lu Sheng. I currently work as an embodied AI algorithm intern at AgiBot and previously worked at the Beijing Academy of Artificial Intelligence (BAAI).
My research interests include embodied intelligence, vision-language-action (VLA) models, robotic spatial reasoning, active perception, and long-horizon manipulation. I am particularly interested in how robots build and use visual memory to acquire information, reason in 3D space, and act reliably in real environments.
lyb371568753I am open to research collaboration on embodied intelligence, VLA models, active perception, and robot learning. Please feel free to reach out by email.
News
- 2026.08: Won 1st place in fully autonomous manipulation at the 2nd World Humanoid Robot Games, hotel guest-service track.
- 2026.07: Our work RoboTracer was accepted to ECCV 2026. See you in Sweden!
- 2026.06: Joined AgiBot as an embodied AI algorithm intern, focusing on VLA training and real-robot deployment.
- 2026.01: Released the RoboBrain 2.5 technical report on depth-aware 3D spatial reasoning and temporal value estimation.
- 2025.06: Joined BAAI as an embodied AI algorithm intern.
Publications & Reports
RoboBrain 2.5: Depth in Sight, Time in Mind
Technical Report, arXiv:2601.14352 · 2026
- Contributed to data and model development for depth-aware 3D spatial reasoning and dense temporal value estimation, improving spatial and temporal modeling in embodied foundation models.
ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation
2026 · First author
Overview: ActiveArena is a unified testbed for studying how robots actively acquire, maintain, and use information during manipulation. It introduces ActiveArena-Sim, 35 tasks across five categories, rich memory annotations, and ID/OOD protocols with disjoint scenes and unseen distractors.
My contribution: Proposed the benchmark and its evaluation protocol; built ActiveArena-Sim and ActiveArena-Bench; developed 13 VLA configurations to analyze memory writing, memory capacity, proprioception, subtask supervision, and high-level planning.
Learning Active Perception and Manipulation via Spatio-temporal Visual Memory
2026 · Co-first author
Overview: ActiveZero formulates active perception as information-driven spatio-temporal memory management. A unified VLA model expands memory through exploration, retrieves relevant evidence for action, and filters memory online for efficient long-horizon interaction.
My contribution: Built the end-to-end active-perception VLA and its memory expansion, retrieval, and filtering mechanisms; contributed to ActiveMem and ActiveBench, including large-scale training data, VQA evaluation, and long-horizon simulated manipulation.
RoboTracer: Mastering Spatial Trace with Reasoning in Vision-Language Models for Robotics
ECCV 2026 · 2025–2026 · Co-first author
Overview: RoboTracer is a 3D-aware vision-language model for spatial tracing, combining universal spatial encoding, metric-aware supervised fine-tuning, and metric-sensitive reinforcement fine-tuning. It introduces TraceSpatial, a 30M-pair dataset, and TraceSpatial-Bench for multi-step metric-grounded reasoning.
My contribution: Contributed to large-scale robot manipulation and spatial reasoning data construction and cleaning across AgiBot, DROID, and RoboTwin; designed the unified pipeline for trajectory extraction, spatial consistency checks, and anomaly filtering.
Education
Research focus: embodied intelligence, VLA models, robotic spatial reasoning, active perception, and memory.
Research & Industry Experience
- Develop embodied policies for humanoid robots, covering VLA training, active perception, task planning, and long-horizon manipulation.
- Participate in visual perception, system integration, and real-world reliability validation for complex service tasks.
- Worked on 3D spatial reasoning, robot trajectory generation, spatial data construction, and embodied-model evaluation.
- Built data-processing pipelines, simulation benchmarks, and evaluation systems over real and simulated robot data.
Competition & Honors
- 2nd World Humanoid Robot Games · Hotel guest service (Aug. 2026): 1st place in fully autonomous manipulation. Contributed to the team and to the deployment of long-horizon service tasks, including luggage handling, room restocking, and room organization.
- National Undergraduate Mathematics Competition: First Prize (2021, 2023).
- National Undergraduate Physics Competition: First Prize (2021, 2022).
- National Physics Olympiad for Secondary School Students: Provincial First Prize (36th competition).
- Beihang University scholarships: Competition Special Scholarship (2021–2022); First-class Academic Scholarship (2022–2025).
- Outstanding Student: Beihang University (2020–2021, 2023–2024).
Skills
Embodied AI
VLA training, active perception, visual memory, spatial reasoning, long-horizon manipulation, and real-robot deployment.
Data & Evaluation
Robot trajectory processing, dataset construction, benchmark design, simulation evaluation, and error analysis.
Engineering
Python, PyTorch, C/C++, Java; distributed data pipelines and multimodal system integration.
Teaching & Service
Teaching assistant for Linear Algebra, Data Structures, and Computer Vision; Deputy Minister of the Beihang Student Science Association Publicity Department.