Engineer Jobs

About The Position

Requirements

• Strong background in machine learning research, with practical experience training large models such as VLMs, LLMs, or robot policies.
• Hands-on experience with pre-training data pipelines, synthetic data generation, or large-scale multimodal training.
• Familiarity with fine-tuning and alignment techniques, including LoRA, RLHF, DPO, and instruction tuning.
• Proficiency in PyTorch.
• Experience with NVIDIA Isaac Sim, MuJoCo, or physics-based simulation engines.
• Ability to design evaluation frameworks that measure task-specific real-world performance beyond standard benchmarks.
• Systems thinking and understanding that training data quality and domain alignment are especially important in constrained physical applications.

Nice To Haves

• PhD or graduate degree in machine learning, robotics, or a related field.
• Experience with embodied AI, robot learning, or sim-to-real transfer.
• Background in industrial or manufacturing domains.
• Experience with distributed training on GPU clusters, including SLURM, DeepSpeed, or FSDP.

Responsibilities

• Design and manage synthetic data generation pipelines using physics simulation, including domain randomization, procedural scene generation, and sim-to-real transfer validation.
• Curate and annotate manufacturing training data, including CAD data, process plans, equipment programs, assembly sequences, sensor streams, and quality inspection records.
• Fine-tune and adapt pre-trained VLMs, VLAs, and LLMs for manufacturing-specific tasks such as DFM co-pilot, fault detection, assembly instruction generation, and robotic policy learning.
• Implement post-training alignment methods such as RLHF, DPO, or preference optimization to align model behavior with manufacturing task success criteria.
• Build domain-specific evaluation benchmarks for manipulation accuracy, instruction-following fidelity, process compliance, and failure mode detection.
• Collaborate with the VLA Research Scientist and AI Infrastructure Engineer to close the training loop between simulation, real factory data, and deployed models.