Researcher, Locomotion

SECOND TALENT SG PTE. LTD.; Singapore; On-site; SGD 100,000 - 180,000; Full time

Job summary

SECOND TALENT SG PTE. LTD. is seeking an experienced reinforcement learning engineer to push legged locomotion research from simulation to real hardware in Singapore. You will own the policy design, training, and deployment for robust bipedal and whole-body locomotion on real robots.

Collaborating with hardware, controls, and robotics researchers, you will refine sim-to-real pipelines, design domain randomization, and publish open-source findings to advance the community.

Qualifications

• Deep hands-on experience with reinforcement learning for continuous control, legged locomotion, or whole-body control.
• Proven track record of deploying learned policies onto real robots.
• Strong command of a physics simulator such as MuJoCo or Isaac, including reward shaping and domain randomization.
• Fluency in Python and modern RL tooling, with comfort working in a ROS2-based control stack.
• A bias for shipping - prioritizing real hardware iteration over polishing results in simulation.
• Strong diagnostic thinking about why a policy fails, not just whether it does.

Responsibilities

• Design, train and ship reinforcement learning policies for bipedal and whole-body locomotion on real hardware.
• Own the sim-to-real pipeline end to end, from training environment to first hardware run.
• Push balance, gait and recovery behavior past demo stage into something that survives pushes, slips and untrained terrain.
• Build and refine reward design, domain randomization and training environments in a physics simulator (e.g. MuJoCo).
• Close the loop between simulation and hardware using real telemetry, feeding failures back into the next policy iteration.

Skills

Reinforcement learning

Continuous control

Deploying to real robots

MuJoCo

Isaac

Python

ROS2

Domain randomization

Reward shaping

Policy iteration

Tools

MuJoCo

Isaac

ROS2

Job description

Our client is an applied R&D lab building an open-source humanoid robot platform and the full software stack behind it, working across hardware, locomotion, autonomy, simulation and infrastructure. This role owns the policies that take the robot from walking to running, recovering and moving robustly through the real world - not just performing well in simulation.

What You'll Do

• Design, train and ship reinforcement learning policies for bipedal and whole-body locomotion on real hardware
• Own the sim-to-real pipeline end to end, from training environment to first hardware run
• Push balance, gait and recovery behavior past demo stage into something that survives pushes, slips and untrained terrain
• Build and refine reward design, domain randomization and training environments in a physics simulator (e.g. MuJoCo)
• Close the loop between simulation and hardware using real telemetry, feeding failures back into the next policy iteration
• Collaborate closely with hardware, controls and manipulation researchers, since whole-body control spans team boundaries
• Publish or open-source findings and tooling so the broader community can build on the work

Core Key Skills

• Deep hands-on experience with reinforcement learning for continuous control, legged locomotion, or whole-body control
• Proven track record of deploying learned policies onto real robots, not only into papers or simulators
• Strong command of a physics simulator such as MuJoCo or Isaac, including reward shaping and domain randomization
• Fluency in Python and modern RL tooling, with comfort working in a ROS2-based control stack
• A bias for shipping - prioritizing real hardware iteration over polishing results in simulation
• Strong diagnostic thinking about why a policy fails, not just whether it does

Nice to Have

• Published work in locomotion, legged robotics, or sim-to-real transfer
• Experience with model predictive control or classical locomotion methods alongside learning-based approaches
• Contributions to open-source robotics or RL projects
• Experience bringing up new hardware and debugging the gap between model and motor