- Salary
- $19 – $65/hr
- Location
- Santa Clara, CA
- Workplace
- Onsite
- Type
- Internship
- Department
- US Internships
- Seniority
- Internship
- Source
- Lever
Description
PlusAI is a Physical AI company pioneering AI-based virtual driver software for factory-built autonomous trucks. Headquartered in Silicon Valley with operations in the United States and Europe, Plus was named by Fast Company as one of the World’s Most Innovative Companies. Partners including TRATON GROUP’s Scania, MAN, and International brands, Hyundai Motor Company, Iveco Group, Bosch, and DSV are working with Plus to accelerate the deployment of next-generation autonomous trucks. If you’re ready to make a huge impact and drive the future of autonomy, Plus is looking for talented individuals to join its fast-growing teams.
This project addresses the sim-to-real gap as in our current simulator pipeline. Closed-loop simulation systems are highly effective at generating massive amounts of diverse, high-fidelity ground truth data. However, they assume perfect perception. However, this ideal GT data does not perfectly align with the BEV feature representations extracted from real-world, noisy sensor data.
The project provides two outcomes:
Bridge Sim-to-Real Gap in Data Augmentation with Closed-Loop Simulation: By translating simulation GTs into a realistic BEV embedding space, the synthetic data becomes directly usable for training and fine-tuning BEV-based planning models.
Allow BEV Space Planning Model to be directly fine-tunable via Self-play RL: Once the sim-to-real gap is bridged in the BEV feature space, the planning models can thus be directly finetuned / post-trained in a self-play RL setting.
The project provides two outcomes:
Bridge Sim-to-Real Gap in Data Augmentation with Closed-Loop Simulation: By translating simulation GTs into a realistic BEV embedding space, the synthetic data becomes directly usable for training and fine-tuning BEV-based planning models.
Allow BEV Space Planning Model to be directly fine-tunable via Self-play RL: Once the sim-to-real gap is bridged in the BEV feature space, the planning models can thus be directly finetuned / post-trained in a self-play RL setting.
Skills
PythonMachine LearningDeep LearningComputer VisionPyTorch
