Research Scientist / Engineer - Post-training & Robot Learning

We're looking for Research Scientists and Research Engineers with deep robotics or autonomous systems domain knowledge to adapt our web-pretrained video model to real robot tasks. Post-training at Rhoda means taking a causal video generation model pretrained on internet-scale data and fine-tuning it on robot-collected demonstrations to produce reliable, generalizable behavior — with as little task-specific data as possible. We hire across levels — from senior to staff. What You'll Do Design and implement RL training pipelines to improve robot policy performance beyond what imitation learning alone achieves — reward design, online data collection, and policy optimization. Develop and apply RL algorithms (PPO, GRPO, or similar) adapted to the video prediction setting, including reward modeling and feedback collection strategies for physical task performance. Design and implement broader post-training pipelines: supervised fine-tuning, preference optimization, and behavioral alignment on robot-collected demonstration data. Work on the inverse dynamics model that translates video predictions into executable robot actions. Build evaluation frameworks for post-trained policies: task success, generalization to novel objects and environments, and failure mode analysis on real hardware. Research methods to efficiently adapt models to new tasks with minimal demonstration data, including in-context generalization and few-shot adaptation. Identify failure modes and systematic weaknesses in deployed robot policies and drive targeted improvements. Iterate quickly between simulation and real robot evaluation to close the feedback loop. Collaborate with the pre-training team to surface what capabilities are missing from the base model and need to be addressed upstream. What We're Looking For Hands‑on experience with robot systems, robotic policy learning, or autonomous systems in an industry or research setting (robotics, self‑driving, or similar physical AI domains). Strong understanding of robot policy learning: imitation learning, behavior cloning, and how RL builds on top of it. Practical familiarity with real robot hardware, deployment constraints, and sensor modalities (vision, proprioception). Solid ML skills with hands‑on PyTorch experience. Ability to diagnose policy failures, reason about distribution shift, and iterate effectively on data and training strategies. Comfort with ambiguity and fast‑changing research priorities. Staff‑level candidates are expected to define technical direction and drive research strategy independently; senior candidates execute complex projects with strong fundamentals and growing scope. Nice to Have (But Not Required) Hands‑on experience with reinforcement learning — reward design, policy optimization, and online RL training loops — applied to real or near‑real environments (robotics, games, simulated physics, or similar); this is a significant plus. Prior industry experience in robotics, autonomous driving, or physical AI (e.g., manipulation, mobile robotics, self‑driving stacks). Experience with teleoperation systems or robot demonstration collection at scale. Familiarity with robot middleware (ROS/ROS2) and real‑time control systems. Experience with simulation environments for robotics (MuJoCo, Isaac Sim, Genesis). Understanding of video generation models and how they connect to action prediction. PhD in Robotics, ML, or a related field. Publication record at ICRA, CoRL, RSS, NeurIPS, or related venues. Why This Role Your work is what makes our robots actually perform tasks reliably in the real world — the direct connection between pre‑trained capability and deployed behavior. Work at a rare intersection: state‑of‑the‑art video generation models applied to real robot hardware, not simulation. Fast feedback loop between model changes and real robot performance. High ownership on a small team where robotics domain expertise is core to the mission. #J-18808-Ljbffr Rhoda ai