Jagannath Prasad Sahoo — AI & Robotics

Current work and interests

My work centres on bipedal robot design and learning-based locomotion: designing small-scale bipeds from scratch, modeling their kinematics and stability, and training RL policies that walk, run and navigate around obstacles. I also work on autonomous multi-agent systems, including heterogeneous UAV–UGV teams, potential-field planners for industrial navigation and vision-based field robots. I’m keen to collaborate on:

Legged & bipedal robotics
Reinforcement learning
Sim-to-real transfer
Multi-agent coordination
Vision Language Models & agentic AI
Motion & path planning
Mechanical design & hardware integration

News & Highlights

Aug 2026Role

Joined AGIHF as a Robotics Engineer Intern

From August to November 2026 I am working on surgical robotics at the Assam Government IIT-G Healthcare Foundation, building a manipulator that holds and positions the surgical scope.
Sep 2026Degree

M.Tech (Research) in AI and Robotics awarded by IIT Mandi

Degree awarded on 02 September 2026 (CGPA 8.57/10). Thesis: “Design, Development and Experimental Validation of a Bipedal Robot for Adaptive Locomotion and Navigation”, supervised by Dr. Amit Shukla.
2026Paper

Five papers accepted: three at ICC 2026, one at IEEE ICARCV 2026 and one at NEECON 2026

At the Indian Control Conference (ICC) 2026: vision-based UAV tracking with deep RL, adaptable bipedal locomotion for perturbation recovery and uneven terrain, and dual-arm cooperative manipulation through shared vision. At IEEE ICARCV 2026 (Halong Bay, Vietnam): hardware model-mediated teleoperation (contributed oral). At NEECON 2026: autonomous buried power cable tracing with EM locator feedback and GPS mapping.
2026Paper

Papers at IEEE ICARA 2026 and ICVGIP

Heterogeneous multi-agent coordination with adaptive perimeter-field navigation (ICARA), and TongueSight, an AI-powered tongue analysis system for digitizing traditional medicine diagnostics (ICVGIP).
2026Award

Outstanding Teaching Assistant Award, IIT Mandi

Recognised for my work as a Teaching Assistant at the Centre for AI & Robotics, IIT Mandi.
Jan 2026Teaching

Teaching Assistant for AR-525 Reinforcement Learning for Robotics

Covered policy gradient and actor-critic methods, built simulation-based assignments and mentored student projects at IIT Mandi. Later received the Outstanding Teaching Assistant Award.
Dec 2025Talk

Presented at the Indian Control Conference (ICC-11)

“Learning multi-skill locomotion in underactuated biped: a waypoint-based reward shaping approach.”

Show earlier highlights Hide earlier highlights

2025Paper

Two papers presented at ICCAS 2025, Korea

Design, kinematic modeling and stability verification of a small-scale biped robot, and kinematics-based real-time distance monitoring for safe multi-arm robot operations.
2025Event

Participant at ROSCon India 2025
2025Paper

Survey presented at IEEE ICCCNT 2025

“From classical control to reinforcement learning: trends in bipedal robot locomotion.”
May 2025Hackathon

TongueTales at the AITech Hackathon, IIT Mandi

Our team built an AI pipeline that analyses tongue images (coating, cracks, papillae, redness) to give explainable health insights. The work grew into the TongueSight paper at ICVGIP.
2023Award

Best Outgoing Student Award, Ramaiah Institute of Technology

Awarded for excellence in academics and co-curricular activities on graduating with a B.E. in Mechanical Engineering (CGPA 8.90/10).
2022Award

Best Paper Award at the 73rd International Astronautical Congress, Paris

First-generation biped robot — 1st-generation biped, designed and built at IIT Mandi

Second-generation 10-DOF biped in the lab — 2nd-generation 10-DOF biped in the lab

Second-generation biped during gantry-supported walking tests — Gantry-supported walking trials

UAV and AGV tracking setup outside the IIT Mandi building — UAV–AGV tracking experiments

Robotics Projects

AdaTerrainDiff: Terrain-Conditioned Adaptive Action Refinement for Unified Bipedal Locomotion

Ongoing work · 2nd-generation biped · IIT Mandi

Diffusion-based locomotion controllers such as BiRoDiff and DiffuseLoco use one mixed latent and a fixed number of denoising steps, so they treat every terrain the same. AdaTerrainDiff splits the encoder into a 128-D state latent and a 16-D terrain latent built from the height map. The terrain latent then sets the action generation: 5 denoising steps on flat ground and up to 10 on stairs, with stochasticity scaled to the terrain. An x₀-prediction denoiser stays stable with only a few steps, and the policy is trained with online PPO.

One policy walks on flat ground, slopes, rough terrain, discrete obstacles and stairs. Averaged over all terrains it reaches 81.0% success, versus 78.0% for BiRoDiff and 32.1% for DiffuseLoco. The biggest gains are on discrete terrain (+12 points) and steep slopes (+5.2 points). On the 10-DOF hardware biped, sim-to-real trials completed 98.3% of each 60-second episode, with cost of transport within 13% of simulation.

Diffusion policy
Terrain-aware RL
PPO
Isaac Gym
Sim-to-real

Terrain-Adaptive Locomotion for a 10-DOF Biped with Shared Autonomy

2nd-generation biped · IIT Mandi

Industrial inspection sites such as manufacturing floors, power plants and refineries have ramps, grated flooring and irregular ground, and people work close by. Most humanoid locomotion work is validated on idealized flat surfaces. This project builds a 10-DOF biped, based on the Berkeley Humanoid Lite URDF, that is meant for those environments.

A reinforcement learning architecture with structured terrain encoding gives stable walking on flat ground, stair ascent and descent, and uneven terrain. Policies are trained in physics simulation with domain randomization for sim-to-real transfer. A low-level safety control loop enforces stability margins and joint-level compliance, so the robot stays predictable and interruptible around human inspectors. The hardware work included custom BLDC actuator control and PID tuning, a range-of-motion study against human joint limits, and gantry-supported sim-to-real trials, including push-recovery tests where the robot is shoved mid-stance and has to regain balance.

Humanoid locomotion
Terrain-aware RL
Shared autonomy
Domain randomization
Push recovery
BLDC actuators

CAD model of the first-generation biped with dimensions and joint layout

Multi-Skill Bipedal Robot with Learning-Based Control

Mar 2024 – Sep 2026 · M.Tech thesis, IIT Mandi

Designed a small-scale biped robot from the ground up, covering mechanical design, kinematic modeling and stability analysis. Trained locomotion policies with PPO, TD3 and SAC that achieve robust walking and running at over 90% success, using waypoint-based reward shaping to learn multiple skills on an underactuated platform.

Policies were developed in Gazebo, PyBullet and Isaac Gym and deployed on hardware through a ROS2 sim-to-real pipeline.

Reinforcement learning
Bipedal locomotion
ROS2
Isaac Gym
Sim-to-real

Success rate, path efficiency, failure modes and energy efficiency across obstacle densities

Hierarchical RL for Collision-Free Locomotion of an Underactuated Biped

Feb 2026 – Jul 2026

A two-level control architecture for obstacle-aware bipedal locomotion. A high-level waypoint policy plans a path through the obstacles, and a low-level gait policy turns it into stable footsteps. The policy was trained in PyBullet with up to 20 obstacles, then deployed on the 1st-generation biped.

In simulation it reaches 98% success with path efficiency (SPL) of 0.955. On hardware it keeps 92% success and SPL 0.884, a sim-to-real gap of about 6–7%.

Hierarchical RL
Path planning
Obstacle avoidance
PyBullet
Sim-to-real

Reward convergence of DQN, SAC, TQC, TD3 and TD7

Deep RL for Vision-Based UAV Tracking by an Autonomous Ground Vehicle

Jan 2026 – Aug 2026 · Centre for AI & Robotics, IIT Mandi · Accepted at ICC 2026

Persistent tracking of a DJI Matrice 600 UAV by a ground vehicle (Diadem AGV) using deep reinforcement learning, with body-frame velocity commands as actions. A YOLOv8n detector, trained on a custom dataset of 5,000 annotated frames, runs in real time on an upward-facing camera. The centroid of the detected bounding box gives the planar tracking error on the image plane.

A composite reward adds a bounded exponential proximity kernel, a potential-based progress term that is guaranteed not to change the optimal policy, and an action-smoothness penalty. Five algorithms (DQN, TD3, SAC, TQC and TD7) were benchmarked on circular and square trajectories. TD7 reached the lowest steady-state error and converged fastest, and an ablation showed that the reward shaping speeds up convergence for every algorithm.

Website →

Deep RL
YOLOv8
UAV–AGV teams
Visual servoing
Reward shaping

Autonomous Buried Power Cable Tracing

Research Associate work · IIT Mandi · Accepted at NEECON 2026

A ground robot that follows buried power cables using electromagnetic locator feedback. It steers to keep the signal strong, avoids obstacles without a predefined goal, and builds a GPS map of the cable route as it drives. The first video shows hardware validation outdoors: the robot tracing a buried cable with an EM locator, seen from overhead, from the ground and on the locator screen (5x speed). The second shows the ROS/Gazebo simulation of the same task. This is part of my work as a Research Associate at IIT Mandi.

Field robotics
Hardware validation
ROS / Gazebo
Sensor fusion
GPS mapping

BLDC Actuator Control & Tuning

Characterization and PID tuning of the BLDC joint actuators that drive the 2nd-generation biped.

Multi-Skill Locomotion (ICC-11)

SAC, TD3 and DDPG compared on push recovery, walking, uneven terrain and stair descent with waypoint-based reward shaping.

TongueTales / TongueSight

AI tongue-image analysis (coating, cracks, papillae, redness) with explainable scores. Built at the AITech Hackathon, IIT Mandi (2025), then published at ICVGIP.

MPPF: Mobile Probe Potential Field

Probe-based potential field planner for real-time multi-agent navigation in industrial environments.

Model-Mediated Teleoperation

Teleoperation of a force-blind, low-cost manipulator using a hardware model and a single-scalar contact sensor.

Smart Water Monitoring, IISc

IoT-based flow monitoring with anomaly detection, deployed at 98% accuracy.

Turbofan Predictive Maintenance

Ensemble ML models for remaining-useful-life prediction of aircraft engines at 92% accuracy.

Projects in Mechanical & Aerospace Engineering

Double Diaphragm Pump Development — Dover India

New product development, R&D and automation for double diaphragm pump systems. FEA and stress testing, plus a 15% cost reduction through VAVE, design optimization and DFM.

6-DOF Vibration Isolation System for Spacecraft — ISRO

Designed and analyzed a six degree-of-freedom vibration isolation system. Validated natural frequencies against analytical models and ran vibration testing from 20 to 2000 Hz.

STARDUST RIT Rocketry & Spacecraft Propulsion

Student coordinator for Aerodynamics and Structures (2020–2022). Designed a green bipropellant (N₂O/C₂H₄) propulsion system for CubeSats and a propulsion system for an interplanetary mission to Venus.

Publications

Citations and the full list are on Google Scholar.

Pant, K. R., Sahoo, J. P., Mallela, S., Surya Prakash S. K., and Shukla, A. “Deep reinforcement learning for vision-based UAV tracking by an autonomous ground vehicle.” Indian Control Conference (ICC) 2026. Accepted. [Project website]
Prasad, P., Choudhury, A., Sahoo, J. P., Kant, R., and Shukla, A. “Adaptable bipedal locomotion for perturbation recovery and uneven terrain navigation.” Indian Control Conference (ICC) 2026. Accepted.
Surya Prakash S. K., Gupta, A., Suraj S. K., Anudeep D., Karthik T., Majji, S. T., Sahoo, J. P., Datta, S., and Shukla, A. “Dual-arm cooperative manipulation through shared vision.” Indian Control Conference (ICC) 2026. Accepted.
Pant, K. R., Vohra, M., Sahoo, J. P., Gupta, A., Karunamurthy, J. V., and Shukla, A. “Autonomous buried power cable tracing using electromagnetic locator feedback, goal-free obstacle avoidance, and GPS mapping.” International Near East Conference on Electrical & Energy Systems (NEECON) 2026. Accepted.
Datta, S., Sahoo, J. P., Surya Prakash S. K., and Shukla, A. “Hardware model-mediated teleoperation for a force-blind low-cost manipulator with a single-scalar contact sensor.” IEEE ICARCV 2026, Halong Bay, Vietnam. Accepted, contributed oral.
Prajapati, D., Kumar, P., Surya Prakash S. K., Mehra, A., Rana, A., Sahoo, J. P., Reddy, T. J., and Shukla, A. “Reference-coupled heterogeneous multi-agent coordination with real-time adaptive perimeter field navigation.” 12th IEEE ICARA 2026, pp. 307–312.
Sharma, V., Sahoo, J. P., Shukla, A., Ameta, D., Shorey, J., Reddy, S., Shewale, P. P., and Sharma, S. “TongueSight: AI-powered tongue analysis for digitizing traditional medicine diagnostics.” 16th ICVGIP, 2026, Article 22, pp. 1–7.
Sahoo, J. P., et al. “Learning multi-skill locomotion in underactuated biped: a waypoint-based reward shaping approach.” Indian Control Conference (ICC-11), 2025.
Sahoo, J. P., et al. “Design, kinematic modeling, and stability verification of a small-scale biped robot.” ICCAS 2025, Korea.
Surya Prakash S. K., Prajapati, D. K., Narula, B., Sahoo, J., and Shukla, A. “Kinematics-based real-time distance monitoring for safe multi-arm robot operations.” ICCAS 2025, Incheon, Korea, pp. 1491–1497.
Sahoo, J. P., et al. “From classical control to reinforcement learning: trends in bipedal robot locomotion.” IEEE ICCCNT 2025.
Madhu Kumar, T. M., Jyothi Kini, H., Praveen, M., Mohan Kumar, Sahoo, J. P., Deepayan, N. C., and Sindhushree, K. L. “Polyethene composites for X-band electromagnetic interference shielding using conductive MWCNT and graphene nanoplatelets with soft magnetic MgFe₂O₄ nanoparticles.” Functional Composites and Structures, 7(3), 035011, 2025.
Sahoo, J. P., et al. “A novel approach for payload design to streamline deep space missions: molecular nanoscience.” 73rd IAC, Paris, France, 2022. Best Paper Award
Sahoo, J. P., et al. “Design of propulsion system for CubeSat: N₂O and C₂H₄ as green bipropellant.” Journal of Aerospace Sciences and Technologies, 74(2), 2022.
Sahoo, J. P., et al. “Analysis and design of a propulsion system for an interplanetary mission to Venus.” International Research Journal of Engineering and Technology, 9(2), 2022.

Work Experience

Robotics Engineer Intern

Aug 2026 – Nov 2026

Assam Government IIT-G Healthcare Foundation (AGIHF)

Working on surgical robotics: designing and building a robotic manipulator that holds and positions the surgical scope (scope holder).

Project Associate I

Jan 2025 – Jul 2026

Indian Institute of Technology Mandi

Led development of an autonomous pipe inspection system based on computer vision and robotics.
Built the end-to-end pipeline: hardware integration, sensor fusion, path planning and deployment with real-time navigation, obstacle avoidance and defect detection.

Teaching Assistant

Jan 2026 – Jul 2026

Indian Institute of Technology Mandi

AR-525 Reinforcement Learning for Robotics: policy gradient and actor-critic methods, simulation-based assignments and project mentoring.
IC-140 Foundation for Design Practicum: hands-on mechatronic design, prototyping and project evaluation.

Design Engineer

Jul 2023 – Mar 2024

Dover India Private Limited

Led new product development, R&D and automation projects for double diaphragm pump systems.
Conducted FEA and stress testing, and reduced cost by 15% through VAVE, design optimization and DFM.

Project Trainee

Dec 2021 – Dec 2022

Indian Space Research Organisation (ISRO)

Designed and analyzed a six degree-of-freedom vibration isolation system for spacecraft.
Validated natural frequencies against analytical models and performed vibration testing from 20 to 2000 Hz.

Education

M.Tech (Research), AI and Robotics

Mar 2024 – Sep 2026

Indian Institute of Technology (IIT) Mandi · Supervisor: Dr. Amit Shukla · CGPA 8.57/10

Thesis: Design, Development and Experimental Validation of a Bipedal Robot for Adaptive Locomotion and Navigation.
Coursework: Robot Kinematics, Dynamics and Control · Mechatronics · Robot Programming · Modeling and Simulation · Principles of Robot Autonomy · Deep Learning and Applications.

B.E., Mechanical Engineering

Aug 2019 – Jun 2023

Ramaiah Institute of Technology, Bengaluru · CGPA 8.90/10

Best Outgoing Student Award for Academics and Co-curricular Activities.
Coursework: Robotics · Nanotechnology · Machine Design · Kinematics and Dynamics of Machines · Manufacturing Processes · CAD and CAM.

Technical Skills

Robotics & Control

ROS
ROS2
Bipedal robotics
Control systems
Motion & path planning
Kinematics & dynamics

AI & Machine Learning

Reinforcement learning
Deep learning
Computer vision
Vision Language Models
Agentic AI

Simulation & Programming

Gazebo
PyBullet
Isaac Gym
MuJoCo
MATLAB
Python
C / C++
Git
Linux

CAD & Analysis

SolidWorks
UG NX
ANSYS (FEA)
NASTRAN
DFMEA
DFM
NPD · R&D · VAVE

Awards, Leadership & Service

Outstanding Teaching Assistant Award, IIT Mandi
Best Outgoing Student Award, Ramaiah Institute of Technology, 2023
Best Paper Award, International Astronautical Congress (IAC), Paris, 2022
Mentor, Drone and Robotics Bootcamp Group, IIT Mandi (Jun 2025 – Aug 2026): hands-on sessions on drone assembly, ROS and robot control for hundreds of students
Evaluator, ZenTej Season 3 Hackathon, IIT Mandi (Oct 2025): assessed prototypes in a 24-hour sprint on soundness, innovation and feasibility
Student Coordinator, Aerodynamics and Structures, STARDUST RIT Rocketry, 2020 – 2022
NSS Coordinator, organized blood donation camps
Paper presenter at ICC-11 2025, ICCAS 2025 (Korea) and IEEE ICCCNT 2025; participant at ROSCon India 2025