Building OmniBot, an open-source mobile manipulator, at Ohho Robotics

Open to robot learning, AI engineering and ML/AI performance roles

Website · LinkedIn · X · Substack · YouTube

I build the software platform that makes robot learning work (teleoperation data collection, episodic data pipelines, training and policy-serving infrastructure) and I make AI models run fast on real compute: GPU and edge-GPU inference, quantisation and model performance optimisation.

• 🔭 Working on: OhhO Robotics (Founder): the open-source software stack for robot learning on mobile manipulators. Ohho OS is open source (Apache-2.0).
• 🤖 Robot learning platform
  • Teleoperation data collection: VR / mobile teleop pipelines for recording demonstrations.
  • Data pipelines: LeRobot-format episodic datasets, shared camera schemas, bag conversion and episode quality filtering.
  • Training & serving: imitation-learning training infrastructure and a ROS 2 policy-inference node that runs trained policies on the robot.
• ⚡ AI optimisation on compute
  • GPU / edge-GPU inference: TensorRT FP16 engine for visuomotor-policy vision encoders (OmniBotPro).
  • Model compression & benchmarking: SAM 2 post-training quantisation and structured pruning, benchmarked against SegFormer on an RTX 4070 Mobile (SAM2forAV).
  • CUDA & ARM: libfreenect2 fork with CUDA 13.0 fixes and Raspberry Pi 5 / ARM64 support; CUDA programming practice; performance trace analysis (Perfetto).
• ⚙️ Core Stack: PyTorch, LeRobot, ROS 2, TensorRT, ONNX, CUDA, Docker, Python, C++.
• 💡 Interests: efficient inference for robot policies and foundation models (including VLAs) on edge GPUs.
• 📫 Email: varun@ohho-robotics.com

• OmniBot #10: arm emergency stop now drops servo torque; the driver holds position on stale commands and clamps joint limits and per-cycle steps.
• OmniBot #13: the LeRobot policy node skips missing or black camera frames, drops stale inference, and clamps arm deltas.
• OmniBot #12: one shared wrist + bird's-eye camera schema across recording, validation, bag conversion and the policy.
• OmniBot #14: episode quality filter CLI for LeRobot datasets that rejects frozen video, camera timestamp skew, joint jumps and too-short episodes.
• OhhO-VR #5: GameCI workflow that runs the Unity edit-mode test suite in CI.
• ohho-os 1.1.2: released to PyPI through a tag-triggered trusted-publishing workflow with install smoke tests on Ubuntu and macOS.

class VarunVaidhiya:
    def __init__(self):
        self.role = "Robot Learning & AI Inference Engineer"
        self.expertise = [
            "Robot learning platforms: teleop data collection to trained policy",
            "Episodic data pipelines & dataset quality (LeRobot format)",
            "Training & policy-serving infrastructure",
            "GPU / edge-GPU inference (TensorRT, ONNX, CUDA)",
            "Model optimisation: quantisation, pruning, benchmarking",
        ]
        self.current_projects = [
            "OhhO Robotics: open-source robot learning software stack",
            "VR / mobile teleoperation data pipelines",
            "Accelerated on-robot policy inference",
        ]
        self.looking_for = {
            "opportunities": "Robot learning, AI engineering & ML/AI performance roles",
            "collaborations": "Robot learning infrastructure, efficient edge inference",
        }

    def say_hi(self):
        print("Let's make robot learning scale and AI run fast!")

#RobotLearning #ImitationLearning #Teleoperation #DataEngineering #MLInfrastructure #GPUInference #EdgeAI #TensorRT #ModelOptimisation #ROS2 #OhhORobotics #PhysicalAI

🌟 Open to collaborations, research opportunities, and interesting projects in AI & Robotics!

_{🛠️ From varunvaidhiya | Last updated: 2026}