Soham Patil

Master’s student, RRC, IIIT Hyderabad

I am a Master’s student at the Robotics Research Center, IIIT Hyderabad, advised by Prof. Spandan Roy.

My research is on learning for contact-rich dexterous manipulation. I am interested in what robots can infer about their physical interaction with objects, such as contact states, geometry and spatial relationships, from RGB cameras and proprioception, and how that information can drive better manipulation policies on real hardware.

This has led to work on learning contact estimation as a pseudo-tactile signal for in-hand object reorientation, and on dexterous grasping policies that remain robust to uncalibrated camera viewpoints, both validated on physical robots. Next, I am interested in learning this information together with the policy, combining it with real touch where a hand has it, and using it to adapt policies on real hardware and across hands with fewer trials.

Always happy to collaborate or chat robotics! Feel free to reach out at my .

Updated

news

Sep 26, 2026 Attending IROS 2026.
Sep 16, 2026 AnyViewDex, on view-invariant dexterous manipulation from RGB observations, is on arXiv · website
Jun 16, 2026 AeroGrab is accepted at IROS 2026. arXiv · website
Jun 16, 2026 NoContactNoWorries is accepted at IROS 2026. arXiv · website

publications

  1. IROS 2026
    NoContactNoWorries: Estimating Contact through Vision and Proprioception for In-Hand Dexterous Manipulation
    Soham Patil, Avirup Das, Sourabh Bhosale, and Spandan Roy
    In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
    arXiv Website Poster

    A transformer fusing RGB-D observations with proprioception estimates binary hand-object contact states, replacing hardware tactile sensors with a learned pseudo-tactile signal. The predicted contacts drive downstream RL for in-hand object reorientation, generalize zero-shot to novel objects, and are validated on a real robot.

  2. Under Review
    AnyViewDex: View-Invariant Dexterous Manipulation from RGB Observations
    Soham Patil*, O. S. Gunjal*, Sourabh Bhosale, A. Chavare, R. S. Hora, and Spandan Roy
    * Equal contribution
    2026
    arXiv Website Code

    Visuomotor dexterous policies are notoriously brittle to camera viewpoint shifts; regularizing a 2D monocular backbone with privileged 3D geometric supervision from simulation resolves this without depth sensors at test time. Zero-shot sim-to-real transfer on an xArm7 with a 16-DoF LEAP Hand achieves 76.7% grasping success across eight unseen objects and six uncalibrated viewpoints.

  3. Preview for AeroGrab: A Unified Framework for Aerial Grasping in Cluttered Environments
    IROS 2026
    AeroGrab: A Unified Framework for Aerial Grasping in Cluttered Environments
    Shivansh Pratap Singh, Naveen Sudheer Nair, Samaksh Ujjawal, Sarthak Mishra, Soham Patil, Rishabh Dev Yadav, and Spandan Roy
    In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
    arXiv Website

    An end-to-end aerial grasping pipeline for cluttered environments: a language instruction selects the target, the platform actively repositions for better views, a grasp network proposes 6-DoF candidates per view, and collision-aware feasibility checking selects the one to execute. Validated on a real aerial manipulator in cluttered scenes.

stepping stones

DextrAH-RGB on xArm7 + LEAP Hand

Ported the DextrAH-RGB dexterous grasping pipeline to our xArm7 and LEAP Hand, including the geometric fabrics controller, a ros2_control interface for the LEAP Hand, and policy changes for our setup so we can retrain it for our own tasks. Containerised, with Isaac Sim and Gazebo digital twins, and runs the policy at 30 Hz with plenty of headroom.

  • xArm7
  • LEAP Hand
  • Geometric Fabrics
  • ros2_control
  • Isaac Sim
  • Gazebo
  • Docker
  • Reinforcement Learning

Dense Tactile In-Hand Manipulation

Ongoing work on how dexterous policies can use dense binary contact information over the hand, starting with dexterous grasping and also in-hand rotation with the LEAP Hand.

  • LEAP Hand
  • Isaac Lab
  • Reinforcement Learning
  • Tactile Sensing
Simulated hand grasping an object, with the recorded contact points shown

Hand-Object Contact Dataset Generation

Large-scale generation of hand-object contact labels. The first version recorded contacts with Isaac Sim's contact API while the hand actively manipulated objects; this was later updated to better techniques.

  • Isaac Sim
  • NVIDIA Warp
  • Python

LEAP Hand with Custom Tactile Fingertips

Fitted the LEAP Hand's fingertips with FSR sensors that report binary contact, and extended its proprioception-only policy to use these readings. Deployed zero-shot from simulation for in-hand rotation.

  • LEAP Hand
  • FSR Sensors
  • Sim-to-Real
  • Reinforcement Learning
Vision-based tactile sensor mounted in a calibration rig

Vision-Based Tactile Sensor Calibration

Vision-based tactile sensors drift as their gel surface wears over time. Tried out methods for calibrating their shear readings to keep measurements consistent.

  • Tactile Sensing
  • Calibration
  • Computer Vision
Bipedal robot standing in the lab

Bipedal Robot Hardware Bring-Up

Early hardware bring-up of a bipedal robot in our lab, covering low-level motor control and the ros2_control software packages, so that colleagues could build their research on it. The platform was later taken forward by others in the lab.

  • Motor Control
  • ros2_control
  • ROS 2
  • Docker
  • C++

MPCHoloNavigation

Model-predictive-control holonomic navigation in ROS 2, with SLAM Toolbox localisation and Nav2 integration. Course project, Robotics: Planning and Navigation, IIIT Hyderabad.

  • ROS 2
  • MPC
  • SLAM Toolbox
  • Nav2

LiDAR_Camera_Calibration

Extrinsic and intrinsic LiDAR-camera calibration from rosbag data and checkerboard images, using Zhang's method for intrinsics. Course project, Mobile Robotics, IIIT Hyderabad.

  • ROS 2
  • LiDAR
  • Camera calibration

DockerForROS2Development

Pre-built Dev Container images for ROS 2 (Humble, Iron) and ROS 1 (Noetic), with optional Gazebo, published on GHCR.

  • Docker
  • Dev Containers
  • ROS 2
  • ROS 1
  • Gazebo

experience

Google Summer of Code @ ros2-control Report2025
Open-Source ContributorRemote
  • Hardware-diagnostics support for ros2_control: a standardised health and fault-reporting API, new ROS messages (HardwareStatus, HardwareDeviceStatus), a unified Actuator/Sensor/System interface, and default diagnostic nodes with CLI tooling. Mentors: Dr. Bence Magyar, Sai Kishor Kothakota.
Robotics Research Center, IIIT HyderabadFeb – Jul 2024
Research InternHyderabad, India
  • Built an in-house collaborative robot arm and an end-to-end pipeline for testing custom controllers on hardware, supporting research on adaptive and robust control.

education

Jul 2024 – present
IIIT Hyderabad, Robotics Research Center
MS by Research, Computer ScienceHyderabad, India
CGPA 10.0/10 · Advised by Prof. Spandan Roy
Jul 2019 – Jul 2023
MIT World Peace University (MIT-WPU)
B.Tech, Computer SciencePune, India
CGPA 8.55/10

awards

DD Robocon team with their robots
DD Robocon National winner 20202020 – 2023
India's national stage of ABU Robocon, the Asia-Pacific university robotics championship.
  • Winner (2020), representing India at the international final; ranked 2nd (2023) and 4th (2022) nationally among 150+ teams.
  • Led a team designing and building complete competition robots from scratch each season across mechanical design, custom electronics, firmware, motor control and autonomy; built an in-house integrated actuator matching commercial specs at ~half the cost, a wheeled-odometry module (<4% error), and ROS 2 / Nav2 autonomy on register-level AVR/TM4C firmware.
Watch the robots (2 videos)
2023, ranked 2nd nationally.
2022, ranked 4th nationally.
MATLAB Winners Circle2020
Omni-drive robot modeling and control in MATLAB/Simulink.

teaching

IIIT Hyderabad
Teaching Assistant
NPTEL
Teaching Assistant
RRC Robotics Summer School 2025 Materials
InstructorRRC Robotics Lab, IIIT Hyderabad