Research Scientist: Reinforcement Learning Singapore / Munich, Germany →

MindOn; Singapore; Vor Ort; SGD 120.000 - 180.000; Vollzeit

Zusammenfassung

MindOn is building general-purpose intelligence for robots and seeks a Research Scientist focused on reinforcement learning to develop RL post-training and optimization for robot manipulation models. You will extend pretrained models like VLA and WAM, using task feedback, interaction data, and reward signals to improve reliable task completion in real environments.

The role emphasizes rigorous experiments and close collaboration with control and software teams to deploy validated improvements on

Qualifikationen

• A passion for Embodied AI.
• Proficiency with Python and PyTorch; ability to review and validate code via tests.
• Master’s degree or higher in Computer Science, Robotics, AI, Automation or related field; or equivalent practical experience.
• Knowledge of RL fine-tuning methods for VLA, WAM, and multimodal models and ability to reproduce and improve training workflows.
• Familiarity with OpenVLA, π0, Diffusion Policy and related models; autoregressive, diffusion and flow-matching action generation.

Aufgaben

• Develop RL fine-tuning and post-training methods for VLA, WAM, and multimodal models to improve task success and stability.
• Design task rewards, success criteria, and evaluation metrics for sparse rewards and long-horizon tasks.
• Combine demonstrations, offline data, and online interactions into effective training workflows.
• Build data collection, policy updates, and evaluation pipelines to reduce real-robot interaction costs.
• Analyze failures, distribution shift, and training instability while preserving pretrained capabilities.
• Collaborate with robot control, software, and data teams on deployment and validation.
• Build reproducible experiments and evaluations to measure improvements in task success and generalization.

Kenntnisse

Embodied AI

Python

PyTorch

Debugging

Ausbildung

Master's degree or higher in CS/Robotics/AI/Automation

Tools

OpenVLA

π0

Diffusion Policy

SimpleVLA-RL

ConRFT

VLA-RFT

Jobbeschreibung

MindOn is building general-purpose intelligence for robots. We are looking for a Research Scientist focused on reinforcement learning to develop reinforcement learning post-training and optimization for robot manipulation models.

You will build on pretrained Vision-Language-Action (VLA), World Action Models (WAM), and other multimodal models, using task feedback, interaction data, and reward signals to improve reliable task completion in real environments.

Responsibilities

• Develop reinforcement learning fine-tuning and post-training methods for VLA, WAM, and other multimodal models to improve task success rates and execution stability.
• Design task rewards, success criteria, and evaluation metrics, addressing sparse rewards, long-horizon tasks, and reward bias.
• Combine demonstrations, offline data, and online interactions into effective training workflows for robot manipulation.
• Build iterative interaction data collection, policy update, and evaluation pipelines to reduce real-robot interaction costs and continuously improve performance.
• Analyze failures, distribution shift, and training instability while limiting degradation of pretrained capabilities.
• Collaborate with robot control, software, and data teams on deployment, interaction data collection, and real-robot validation.
• Build reproducible experiments and evaluations to measure improvements in task success, robustness, and generalization.

Qualifications

• A deep and genuine passion for Embodied AI is essential.
• Proficiency with AI coding tools is essential: use them effectively for development, debugging, testing, and troubleshooting; independently review generated code and validate its correctness and reliability through tests and real-world execution.
• A master’s degree or above in Computer Science, Robotics, AI, Automation, or a related field, or equivalent practical experience.
• Familiarity with reinforcement learning fine-tuning methods and representative work for VLA and other multimodal action models, such as SimpleVLA-RL, ConRFT, and VLA-RFT; understanding of their training workflows, reward design, data usage, and applicability, with the ability to reproduce, evaluate, and improve these methods.
• Knowledge of OpenVLA, π0, Diffusion Policy, and related models, including autoregressive, diffusion, and flow-matching action generation and corresponding approaches to reinforcement learning optimization.
• Practical experience fine-tuning pretrained robot policies or multimodal action models with reinforcement learning, designing training plans around task success rates and validating results.
• Familiarity with reward design, success detection, interaction data collection, and combining demonstrations with online data.
• Ability to diagnose training instability, reward hacking, distribution shift, and degradation of pretrained capabilities.
• Proficiency in Python and frameworks such as PyTorch, with experience developing, debugging, and optimizing training pipelines.
• Rigorous experimental skills, including benchmark evaluation, ablations, and failure analysis to establish whether improvements are effective.
• Strong problem-solving skills, independent execution, and cross-functional collaboration.

Preferred Experience

• RL post-training experience with VLA, WAM, or other robot foundation models, with measurable improvements in task success.
• Experience with online learning, policy deployment, and debugging on real robots.
• Familiarity with human intervention data, Human-in-the-Loop learning, and combining demonstrations with reinforcement learning.
• Experience developing visual reward models, success detection models, or learned reward functions.
• Experience applying reinforcement learning to diffusion or flow-matching action models.
• Familiarity with distributed training, parallel data collection, and large-scale experiment management.