Responsibilities
- Track, reproduce, and evaluate the latest open-source World Models (e.g., DreamZero, LingBot-VA, UniSim, Cosmos); analyze the strengths and weaknesses of different World-Action Model architectures and produce technical selection reports
- Fine-tune open-source World Model pre-trained weights to adapt to the team's proprietary robot platform and manipulation task scenarios, enabling the World Model to accurately predict future video frames and physical interaction outcomes of robot operations
- Integrate fine-tuned World Models into the VLA Pipeline: explore practical applications such as World Model as Data Augmentation (generating synthetic training data) and Action Scoring (scoring and filtering candidate action sequences)
- Participate in teleoperation data collection to accumulate high-quality robot manipulation video data for World Model training
- Leverage simulation platforms such as Isaac Sim to generate large-scale, diverse synthetic video data to supplement real-world data
- Build a World Model Evaluation Pipeline covering Prediction Quality metrics (FVD / SSIM / LPIPS) and performance gains on downstream VLA task success rates
- Continuously track frontier advances in World Models (DreamZero WAM paradigm, LingBot-VA MoT architecture, Cosmos physics reasoning, etc.) and share findings with the team regularly
Requirements
- Master's degree or above in Computer Science, Artificial Intelligence, or related fields
- Deep understanding of Video Generation / Prediction; proficiency in at least one of: Diffusion Models, Autoregressive Models, or Flow Matching
- Proficient in PyTorch with end-to-end project experience in video generation or visual model training / fine-tuning
- Strong ability to read research papers and reproduce open-source code; capable of quickly running a new World Model and benchmarking its performance
- Familiarity with fundamental Robotics concepts (State Space, Action Space, Reward); willingness to dive deep into Embodied AI
- Solid mathematical foundation (probability theory, optimization, dynamical systems)
- Highly self-motivated with the ability to maintain strong technical sensitivity in the rapidly evolving World Model landscape
Bonus Points
★ Publications related to Video Generation / Prediction at NeurIPS, ICML, ICLR, CVPR, or equivalent venues ★ Familiarity with Model-Based RL (MBRL) or Model Predictive Control (MPC) ★ Experience with large-scale distributed training (multi-node multi-GPU, DeepSpeed, FSDP) ★ Experience generating training data in simulation environments such as Isaac Sim / Habitat / AI2-THOR ★ Hands-on experience with video generation models such as Sora / SVD / CogVideo / Cosmos ★ Ability to independently design a joint training framework for World Model and VLA Policy, or propose an Imagination-Based Planning solution