Robotics Research Engineer (L4) — Physical AI & Robot Learning

Robotics Research Engineer (L4) - Physical AI & Robot Learning

Hyderabad, India · Full-time · On-site with real robots

Build Intelligence That Holds Up in the Physical World

A robot can follow a programmed trajectory. The harder challenge is helping it work when the world does not match the plan: a connector is slightly misaligned, a component arrives in a different pose, or parts in a bin are partially hidden from view.

Perceptyne builds industrial humanoid robots for physical work such as assembly, manipulation, and packaging. Making these robots more capable requires advances not only in AI models, but also in perception, control, data collection, evaluation, and the software that brings those pieces together on real hardware.

We are looking for a senior, hands-on Physical AI engineer to help advance that work. You will explore promising approaches in robot learning, adapt and integrate them with Perceptyne’s robots, and contribute to the evolution of PR-PhI, our modular robot-intelligence platform. Where it makes sense, you will also help investigate new models tailored to our robots and tasks.

You will report to the founders and begin as a senior individual contributor. You will work alongside our existing AI/ML and robotics engineers, contribute directly to critical code and experiments, and be part of shaping the technical direction as the team grows. This is a role for someone who brings depth and judgment to a collaborative engineering effort-not someone arriving to replace the expertise already here.

What You’ll Actually Do

• Your work will span research, software architecture, experiments, and deployment on physical robots. You will collaborate closely with engineers across AI/ML, manipulation, computer vision, controls, embedded systems, and mechanical design.
• Be part of defining priorities and technical approaches for robot learning and Physical AI.
• Evaluate promising robot-learning and physical-intelligence models on Perceptyne hardware. Identify where they work, where they fail, and what it would take to make them useful in practice.
• Adapt, fine-tune, integrate, and deploy selected models for our robot configurations, sensors, action spaces, and manipulation tasks.
• Compare learned approaches against practical alternatives-including classical computer vision, motion planning, visual servoing, force-based control, and task-specific skills-and use the approach that best serves the task.
• Contribute to the architecture and implementation of PR-PhI, connecting perception, robot state, learned policies, planning, skill execution, monitoring, verification, and recovery.
• Set-up new and strengthen our existing data-collection and experimentation workflows, including teleoperation, synchronized sensor and robot-state logging, demonstrations, interventions, dataset quality, versioning, and reproducibility.
• Design meaningful real-robot evaluations that measure task success, robustness to variation, cycle time, intervention needs, latency, and failure modes-not just offline model metrics.
• Build reliable training and inference pipelines, and work through the practical constraints of serving models on robots. Cloud infrastructure may be used where appropriate, with attention to longer-term deployment on local or on-premises compute.
• Work with the broader robotics team to define safe operating boundaries, failure detection, fallbacks, and appropriate human involvement for experimental and deployed capabilities.
• Read and test relevant research quickly, while helping the team distinguish promising results from approaches that are not yet practical for our hardware or customer tasks.
• Write high-quality code, review designs and experiments with peers, mentor engineers through technical collaboration, and help grow the team’s collective capability.
• Explore proprietary model development where experimental results and real-world needs make a convincing case for it.

What We’re Looking For

You are a strong engineer with experience at the intersection of machine learning and physical robots. You enjoy ambitious research questions, but you are equally comfortable investigating a calibration error, improving a data pipeline, or debugging why a policy fails on hardware.

• 5+ years of relevant experience in robot learning, embodied AI, manipulation, applied ML for robotics, autonomous systems, or a closely related area and 8+ years of overall software / product related work experience. We care more about the depth of your work than an exact year count or previous title.
• Hands-on experience training, fine-tuning, evaluating, or deploying learned behavior on real robots. Simulation and offline research experience are valuable, but are not sufficient on their own for this role.
• Strong Python and modern ML-framework skills, particularly PyTorch or JAX. Ability to write maintainable, performant code; C++ experience is valuable.
• Practical familiarity with relevant approaches such as imitation learning, behavior cloning, diffusion or flow-based policies, vision-language-action models, reinforcement learning, or multimodal perception-action systems.
• Sound robotics fundamentals: coordinate frames, kinematics, calibration, cameras, robot state, action representations, control interfaces, grasping, contact-rich manipulation, latency, and failure recovery.
• Experience with robot-learning data: demonstrations, teleoperation, sensor synchronization, dataset design, quality checks, experiment tracking, and repeatable evaluation.
• Experience with cloud GPU infrastructure, distributed training, model serving, CUDA, TensorRT, ONNX, or edge-compute optimization.
• Strong experimental judgment. You establish baselines, define success criteria, investigate failure modes, and know the difference between a compelling demonstration and a robust capability.
• Systems thinking: you understand how learned models must work alongside perception, planning, controls, safety mechanisms, and real-time robot software.
• Ability to work constructively with experienced engineers across disciplines, share technical judgment clearly, and improve work through code review, mentoring, and collaborative problem-solving.
• Grit, intellectual honesty, and willingness to spend substantial time with physical robots when results do not match expectations.
• Willingness to work full-time on-site in Hyderabad.
• Bachelor’s, Master’s, or PhD in a relevant technical field-or equivalent demonstrated experience.

Nice to Have

These are useful, not mandatory. Please apply if your core experience fits the role.

• Experience with dual-arm manipulation, humanoid robots, dexterous end-effectors, industrial assembly.
• Experience building teleoperation tools, operator interfaces, intervention workflows, or large-scale robot-data pipelines.
• Research publications, open-source contributions, or deployed systems in robotics, embodied AI, computer vision, or controls.
• Experience mentoring engineers or helping establish sound research-engineering practices in a growing team.
• Familiarity with industrial automation, manufacturing workflows, or customer-facing robotics deployments.

How We Work

At Perceptyne, the robot is the referee.

A paper result, a benchmark score, a training curve, or an impressive video is more valuable when it helps a real robot perform a useful task reliably. We value first-principles thinking, rapid experimentation, disciplined measurement, and direct observation of what happens on hardware.

We hold a few deliberate tensions:

• Frontier AI, grounded in real robot and customer constraints.
• Fast experimentation, supported by rigorous evaluation and reproducible engineering.
• Learned intelligence where it helps, and classical perception, planning, control, or deterministic logic where it works better.
• Long-term research ambition, connected to near-term useful capabilities.
• Technical depth, combined with clear communication and practical cross-functional collaboration.

We value grit and dedication. When a policy fails on a real connector, a model is unstable outside its training distribution, a camera-calibration issue invalidates a dataset, or an elegant idea does not survive real-time deployment, you do not merely tune until the demo looks better. You identify the failure mode, establish the right baseline, improve the data or system, update the architecture where necessary, and stay with the problem until there is a credible path forward.

You will thrive here if you want to build the intelligence layer behind robots that do real work-and are excited by the difficult, rewarding path from a promising model to a dependable machine.