DynastGNN: Dynamic Spatio-Temporal Hierarchical Graph Neural Network for Visuotactile 6D Pose Estimation of an In-Hand Object
Snehal Dikhale, et al.
Paper in Progress, 2025
Snehal S. Dikhale
Robotics Researcher @ Honda Research Institute
5 yrs industryM.S. Robotics
The industry has mastered seeing. I want to give robots the ability to truly feel. That means cracking tactile representation and using it to unlock dexterous manipulation.
Right now I'm extending Vision-Language-Action architectures with touch, connecting foundation models all the way down to real hardware.
I'm equal parts researcher and engineer, and I genuinely love both. I call it hardware intuition. Always happy to connect, whether it's research, life, or just geeking out about robots.
Research & Practice
Researcher by Nature, Engineer by Practice
Simulation
High-fidelity physics modeling
Domain randomization
Sim-to-real transfer
Algorithm
Tactile foundation encoders
Tactile VLM/VLA
Spatio-temporal GNNs
Hardware
Real-world deployment
Tactile sensor integration
Multi-fingered manipulation
Experience
Research Engineer II, Robotics
CurrentHonda Research Institute · Apr 2024 – Present
25%
tasksuccess
300k+
multimodalepisodes
5
patentsfiled
- Architected a high-fidelity multimodal MuJoCo simulator for the Allegro hand with force-taxel tactile sensing, GPU-parallelized via NVIDIA Warp; scaled to 300k+ dexterous-task rollouts across 150+ objects
- Designed and pretrained a family of tactile foundation encoders, including a hierarchical taxel-based graph transformer aligned with hand anatomy, powering downstream dexterous manipulation tasks
- Trained TacVLM, a tactile vision-language model for closed-loop dexterous grasp-failure reasoning, supported by an agentic LLM-driven labelling pipeline that produced causal training signal at scale
- Built TacVLA, a multimodal Vision-Language-Action policy fusing optical and taxel-based tactile, vision, and proprioception with a cross-attention DiT action head; deployed on physical multi-fingered hardware for contact-rich peg insertion via teleoperation-collected imitation data; mentored a Fall 2025 PhD intern through the full project lifecycle
- Selected as 1 of 30 individuals across all of Honda North America for leadership development training
Tactile Foundation ModelsVLM/VLAMuJoCo WarpDexterous ManipulationIntern Mentorship
Research Engineer I, Robotics
Honda Research Institute · Sep 2020 – Apr 2024
65%
sim-to-realgap
3×
tactileresolution
100+
robotexperiments
- Built an end-to-end perception pipeline for dexterous in-hand pose estimation under heavy occlusion, owning the full stack from algorithm design to deployment; validated robustness via 100+ real-robot experiments
- Built one of the first tactile simulation setups in Unreal Engine, engineering a 220k-sample domain-randomized visuotactile dataset (RGB-D + tactile) that closed the Sim-to-Real gap by 65%
- Designed a spatio-temporal graph-based representation learning framework combining video, depth, proprioception, and taxel-level tactile data, improving pose estimation temporal consistency by 30% in dynamic environments
- Developed contrastive learning techniques for taxel-based signals, achieving a 3× improvement in tactile resolution to enhance fine-grained manipulation capabilities
- Designed hardware-agnostic CNN and graph representations for vision, depth, and 3D tactile sensor fusion, enabling generalization across multiple robots and reducing position error by ~35% and angular error by ~64% over vision-only baselines (IEEE RAL 2022, 76+ citations)
Sim-to-RealVisuotactile Sensing6D Pose EstimationSpatio-Temporal GNNsContrastive Learning
Graduate Researcher
Worcester Polytechnic Institute · 2018 – 2020
78%
grasp success
3.8
GPA
4
projects
- M.S. Thesis: Built a simulation and benchmarking framework (Gazebo, MoveIt, Panda Arm, RealSense) to evaluate deep learning grasping algorithms; achieved 78% success with GQCNN and 65% with GPD on RGB-D data
- Human-Robot Handover: Designed interaction experiments using ROS/Python; trained ProMPs to predict Object Transfer Point (RMSE < 0.2m)
- Simulation of Control Techniques: Implemented Robust, PD, and PD+Gravity controllers for Baxter Arm; evaluated via MATLAB simulations
- Predicting Building Energy Consumption: Applied regression, random forest, and neural networks; achieved RMSE 1.27, ranked top 30% on Kaggle
- WIN Women's Young Investigator Fellowship – Awarded to the top 4 graduate female researchers in STEM at WPI
- Mentor, Women's Research and Mentorship Program – Mentored 1 undergraduate and 2 high school students; led robotics and 3D printing workshops
ROSGazeboMoveItDeep Learning for Grasping
Selected Patents & Publications
Provisional Patents
- S. Dikhale, et al. Systems and Methods for Embodiment-Aware Representation and Processing of Distributed Sensor Data for Physical Artificial Intelligence · US Provisional Patent, 2026 (filed)
- S. Dikhale, et al. Tactile-conditioned Vision-Language Models for Failure Reasoning in Multi-Fingered Dexterous Manipulation · US Provisional Patent, 2025 (filed)
- A. Shahidzadeh*, S. Dikhale, et al. Context-Aware Multimodal Action Planning Using Tactile, Vision, and Language · US Provisional Patent, 2025 (filed) · * Mentored Intern
Technical Skills
Research & AI
Embodied AIVision-Language Models (VLMs)Large Language Models (LLMs)Multimodal LearningDexterous ManipulationSim-to-Real Transfer3D Tactile RepresentationGraph Neural NetworksTransformersContrastive LearningComputer VisionDeep Learning
Software & Frameworks
PythonC++PyTorchHugging FaceCUDAROSDockerGit
Simulation & Tools
MuJoCoNVIDIA WarpUnreal EngineGazeboMoveItBlender
My Story
Medium
Robotics Chose Me, But I Choose It Every Day
A personal essay on what it really means to build a career in robotics — the doubt, the obsession, the hardware that breaks at 11pm, and why I'd choose it all over again.
Read on Medium