[Remote] Robotics Research Engineer, World Model & Synthetic Data

Note: The job is a remote job and is open to candidates in USA. Miraxis AI builds AI-assisted data generation systems for Physical AI and robotics. They are seeking a Robotics Research Engineer, World Models & Synthetic Data to develop representation, simulation, and transition-prediction layers for their robotics data platform, focusing on egocentric video understanding and synthetic data generation.ResponsibilitiesBuild world-model and transition-evidence pipelines for robotics video and demonstration dataRepresent physical claims as state-before, claimed interaction, expected state-after, observed state-after, evidence, uncertainty, and provenanceIntegrate video representation models such as V-JEPA-style models, VideoMAE, DINO-style encoders, or other frozen foundation models into reproducible inference pipelinesBuild latent residual scoring over video and multimodal representationsDesign lightweight transition predictors, probes, or calibrated residual models over frozen embeddingsBuild digital twin workflows for selected robotic tasks, scenes, objects, environments, and failure modesUse simulation and digital twins to generate controlled variations: object pose, lighting, occlusion, camera viewpoint, clutter, distractors, contact events, action perturbations, and rare failuresEvaluate synthetic and simulated data for usefulness, not just visual realismCompare simulated or synthetic data interventions against real-world review outcomes and downstream evaluation metricsBuild pilot reports on routing and synthetic-data value, including Lift@k, correction capture rate, confidence intervals, calibration, and failure analysisWork with computer vision engineers to incorporate masks, tracks, dense visual features, clip embeddings, object state, and scene structure into the transition-evidence layerWork with platform engineers to ensure outputs are versioned, reproducible, traceable, and stored as auditable artifactsSupport post-pilot model development using human correction traces, model disagreements, deterministic validation failures, simulation deltas, and final accepted annotationsKeep world-model and simulation outputs as evidence, stress-test signals, and data-generation aids. They should not automatically finalize annotations or replace human review for high-risk casesSkillsStrong Python and PyTorch experienceHands-on experience with video models, robotics perception, simulation, synthetic data, self-supervised representation learning, temporal action recognition, action anticipation, or latent predictionExperience working with frozen foundation models and building task-specific probes, heads, residual models, or scoring functions on top of embeddingsStrong understanding of video representation evaluation, uncertainty, calibration, weak signals, and human-in-the-loop MLExperience with simulation or digital twin workflows for robotics, perception, physical scenes, or embodied AIExperience with GPU inference, batch processing, experiment tracking, artifact versioning, and reproducible evaluationAbility to design evaluation plans that avoid overfitting small, biased, or noisy pilot datasetsComfort working with egocentric and robotics video failure modes, including occlusion, hidden hands, camera motion, fast motion, object confusion, temporal aliasing, weak labels, ambiguous contact, and sim-to-real gapsStrong statistical judgment, including practical use of Lift@k, AUROC, AUPRC, calibration curves, Brier score, confidence intervals, and bootstrappingExperience with V-JEPA, JEPA-style architectures, VideoMAE, DINO, DINO-WM, latent world models, action-conditioned prediction, EPIC-KITCHENS, Ego4D, BridgeData, DROID, RoboCasa, ManiSkill, or other robotics video datasetsExperience with digital twins, robotics simulation, or synthetic data generation using tools such as NVIDIA Isaac Sim, Omniverse, MuJoCo, Genesis, Unreal, Unity, Blender, RoboSuite, or similarExperience with domain randomization, procedural scene generation, sim-to-real validation, synthetic data filtering, or counterfactual data generationExperience with Physical AI, robotics perception, embodied datasets, human demonstration data, temporal segmentation, or state-change detectionExperience building active-learning, uncertainty-ranking, review-routing, or human-review prioritization systemsExperience evaluating model outputs against human corrections, operational review outcomes, or downstream robotics performanceExperience with video artifact storage, vector search, MCAP, Rerun, ROS/ROS 2, or similar tools for multimodal data inspectionExperience working with 3D assets, object state, scene graphs, camera calibration, pose estimation, or spatial representationsCompany OverviewMiraxis is building the data layer for physical AI. The next generation of robots and embodied AI systems will not be trained by language alone. It was founded in 2026, and is headquartered in Newark, Delaware, US, with a workforce of 2-10 employees. Its website is https://miraxis.ai.