42dot je podjetje za umetno inteligenco na področju mobilnosti, ki deluje na presečišču avtomobilske tehnologije in logistike. Njegovo delo je osredotočeno na tehnologije vozil, ki jih opredeljuje programska oprema, avtonomna električna vozila ter programsko in podatkovno infrastrukturo, potrebno za razširljiv mestni promet. Podjetje razvija programsko opremo za vozila, podatkovne platforme voznih parkov in modele umetne inteligence za zaznavanje okolice in upravljanje ter TAP!, svoj operacijski sistem za mobilnost, namenjen uvajanju avtonomnih storitev. Z raziskavami in sodelovanjem na področju avtomobilske elektronike ter arhitektur, ki temeljijo predvsem na programski opremi, si 42dot prizadeva, da bi bili sistemi mobilnosti bolj avtonomni, prilagodljivi in preprostejši za upravljanje.
Reinforcement Learning Engineer, Autonomous Driving
Develop and optimize reinforcement learning policies for 42dot’s autonomous-driving models. Build self-play simulations, scenario rollouts, and safety evaluations to improve driving performance.
Opis
- Design observation and action interfaces between VLA policies and simulators, and develop closed-loop reinforcement learning post-training using methods such as GRPO and PPO.
- Build self-play training in vector- or occupancy-based driving simulators, where the ego vehicle and traffic agents interact with current policies and previous checkpoints.
- Develop candidate action and trajectory rollouts across possible future scenarios, then select actions using expected return and safety constraints.
- Evaluate policy safety and generalization with closed-loop benchmarks.
- Analyze reward hacking and simulator bias.
- Lead promising reinforcement learning ideas from design through large-scale training and rigorous evaluation to integration with driving models.
Zahteve
- At least three years of experience in machine learning, reinforcement learning, or robotics, or equivalent research and development capability.
- Experience implementing policy learning algorithms and conducting training and evaluation with Python and PyTorch.
- Understanding of policy optimization, reward design, and credit assignment, with the ability to design reinforcement learning experiments.
- Experience training policies in simulation or multi-agent environments and verifying improvements against baselines.
- English communication skills sufficient to discuss technical issues and collaborate with overseas offices.
- Experience developing self-play, multi-agent reinforcement learning, or policy population training systems (preferred).
- Experience applying reinforcement learning post-training methods such as GRPO or PPO to VLA or generative policies (preferred).
- Experience training policies in vector- or occupancy-based driving simulators, or developing world-model-based rollout planning (preferred).
- Experience with large-scale parallel rollouts, distributed training, or sim-to-real validation (preferred).
Ugodnosti
- Preference is given to veterans and people eligible for employment protection under applicable laws.
- Preference is given to applicants with a registered disability certificate.
- Reference checks may be conducted with the applicant’s consent after the interview process.
- A three-month probationary period may apply.