Jialeng Ni

InternManip: All-in-one Manipulation Learning Framework

Research Intern | OpenRobotLab, Shanghai Artificial Intelligence Lab — Shanghai, China
05/2025–08/2025

• Developed a unified framework spanning manipulation policies (ACT, DP, Pi-0, GR00T-N1/1.5), robot morphologies (Franka, bimanual AgileX Piper arm), training datasets, and benchmarks (GenManip-v1, Calvin, Simpler-Env).
• Integrated language-conditioned Diffusion Policy; standardized evaluation with GR00T-N1.5 and pi-0 VLA baselines.
• Co-organized the IROS 2025 Challenge on Multimodal Robot Learning in InternUtopia and real world.

Challenge (Co-organized) & Repository: IROS 2025 Challenge | GitHub Repository

Touch-Aware Multimodal Robotic Manipulation with Magnetic Tactile Sensors

Bachelor’s Thesis | UM-SJTU Joint Institute, Shanghai Jiao Tong University — Shanghai, China
05/2025–08/2025

• Proposed a modular “6+1” touch-aware manipulation framework that decouples gripper intelligence from 6DoF arm planning, enabling high-frequency tactile feedback while maintaining efficient arm motion planning.
• Designed a wearable visuo-tactile data collection device for intuitive fingertip demonstrations and fused it with on-robot teleoperation data to train a robust, scalable gripper control model.
• Validated on a Realman arm–gripper platform controlled by a deep visuo-tactile model, safely handling fragile and deformable objects (e.g., paper clip, grapes, etc.) with adaptive gripping force.

NeuralMetric: Transformer-Based Safety Metric for Autonomous Driving

Mcity, University of Michigan — Ann Arbor, MI
02/2024–12/2024

• Proposed NeuralMetric, a Transformer-based evaluation model that learns directly from historical trajectories to capture potential collision risk, addressing limitations of heuristic-driven and latency-prone safety metrics.
• Constructed three large-scale simulation datasets with realistic accident distributions; mitigated extreme class imbalance via weighted losses and sampling strategies, significantly improving detection of rare but critical events.
• Validated across highways, intersections, and roundabouts, achieving state-of-the-art predictive accuracy and real-time deployment (2 ms on GPU / 76 ms on CPU, 100–1000× faster than prior methods).

Publications: [1] X. Yan, H. Sun, J. Ni, H. Zhu, S. Feng, and H. Liu, "NeuralMetric: An Accurate and Efficient Real-time Safety Metric for Automated Driving Systems" Under Review at the 26th International Symposium on Transportation and Traffic Theory.

Ablation Study on Tri-Perspective View on Plane Configurations

Robotics, University of Michigan — Ann Arbor, MI
10/2024–12/2024

• Conducted an ablation study on Tri-Perspective View (TPV) for 3D semantic occupancy prediction on Panoptic nuScenes datasets; achieved a 48% mIoU with two-plane setups versus 40% with traditional TPV.
• Demonstrated that two-plane configurations are sufficient for effective vision-based occupancy prediction.

Open-Source Code & Paper: GitHub Repository | Paper | Poster