Build with us.
The next paradigm needs builders.

What You'll Work On

This research role focused on three connected directions:

1. causal discovery from high-dimensional observational and interventional data;
2. causal world models for physical systems;
3. causal foundation models across language, vision, time-series, and multimodal data.

We do not expect candidates to be experts in all three. We are looking for researchers who can make deep progress on one of these directions while helping connect it to the others.

• Develop methods for causal discovery from high-dimensional observational and interventional data, and apply them to real-world data, including pixels, trajectories, sensor systems, biological systems and experimental data
• Build agentic causal discovery loops where models propose hypotheses, choose interventions, call tools or simulations, interpret outcomes, and update causal knowledge base— with scientific discovery as a key long-term direction
• Develop causal world models that represent latent state, mechanisms, interventions, and counterfactual dynamics in physical, scientific, and agentic systems
• Investigate structural limitations of current foundation models for causality: correlational objectives, weak intervention semantics, entangled representations, brittle OOD behavior, and shallow counterfactual reasoning
• Prototype new modeling, training, and evaluation methods that bring causal structure into language, vision, time-series, and multimodal foundation models
• Work with policy, robotics, simulation, and multimodal modeling teams to test whether causal structure improves intervention selection, prediction, planning, or generalization

Background We're Looking For

• PhD in machine learning, statistics, causal inference, or a related field — or equivalent research output
• Strong publication record at NeurIPS, ICML, ICLR, UAI, AISTATS, ACL, CVPR, CoRL, RSS, or comparable top venues
• Research track record in at least one of: causal discovery (PC, GES, NOTEARS, DCDI, or related), causal inference, causal representation learning, world models, AI for science, agentic discovery systems or mechanistic interpretability
• Experience working with nontrivial data beyond clean tabular settings: temporal data, multimodal observations, interaction traces, simulations, robotics data, scientific measurements, or interventional experiments
• Strong implementation ability in PyTorch, JAX, or equivalent frameworks; able to train large models, design ablations, build evaluations, and debug empirical failure modes

Nice to Have

• Experience with physical AI, robotics, or biological systems, or scientific datasets
• Familiarity with interventional experimental design or active learning under causal constraints
• Experience with mechanistic interpretability or causal analysis of trained neural networks

What You'll Work On

• Train policy models for long-horizon physical and agentic tasks using demonstrations, teleoperation, real robot trajectories, simulation rollouts, feedback, and online experience
• Develop policy learning methods across imitation learning, offline RL, online RL, model-based RL, hierarchical RL, diffusion/flow policies, and VLA-style language-conditioned policies
• Build the model-based RL framework based on causal world models, including sample-efficient exploration and planning, learning with imagined rollouts, counterfactual policy evaluation, and failure attribution
• Build evaluation loops that measure task success, robustness, recovery, generalization, safety margins, and regressions across policy versions
• Diagnose policy failures end-to-end: data coverage, reward design, representation error, simulator mismatch, perception failure, control limitation, latency, or hardware constraint

Background We're Looking For

• PhD in reinforcement learning, control theory, or machine learning — or equivalent research depth
• Experience with BC/IL/offline RL/online RL/model-based RL
• Hands-on experience training policies from real robot data or large-scale embodied interaction data.
• Experience training policies in simulation and transferring to physical hardware
• Familiarity with causal inference, causal MDPs, or decision-making under intervention
• Publications at NeurIPS, ICML, ICLR, CoRL, RSS, or equivalent

Nice to Have

• Experience with vision-language-action (VLA) models or language-conditioned policies
• Familiarity with causal credit assignment, structural causal models, or do-calculus
• Experience with sim-to-real transfer and domain randomization

What You'll Work On

• Build embodied foundation models over video, language, actions, robot trajectories, proprioception, depth, tactile/force signals, simulation traces, and agent-environment interaction data.
• Develop architectures for embodied agents: perception, physical scene understanding and reasoning, memory, instruction grounding, planning, action representation, and feedback.
• Design representation learning methods that extract causally meaningful variables from high-dimensional observations: object states, contact relations, latent physical properties
• Develop data mixtures across internet video, robot data, teleoperation, simulation, egocentric video and language annotations.
• Collaborate with the Causal AI team to connect multi-modal representations to downstream causal structure learning and planning

Background We're Looking For

• PhD in machine learning, computer vision, or NLP — or equivalent research output
• Deep familiarity with modern deep learning architectures and self-supervised learning methods
• Strong track record in training substantial video, multimodal, VLA, world-model, time-series, robot, autonomous-system, or embodied-agent models.
• Experience with dataset curation, filtering, data-mixture design, scaling behavior, and training dynamics.
• Evidence that models improved downstream physical scene understanding, embodied reasoning, planning, policy learning, or task performance.
• Publications at NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, or equivalent

Nice to Have

• Familiarity with causal representation learning, disentanglement, or identifiable latent variable models

What You'll Work On

• Build and maintain the training stack for large embodied, multimodal, video, and policy models across GPU clusters
• Design and maintain distributed training systems for large multi-modal models across GPU and TPU clusters
• Build high-throughput, fault-tolerant data pipelines for heterogeneous research data: video, depth, proprioception, force signals, trajectory logs
• Optimize training efficiency: mixed precision, gradient checkpointing, model and data parallelism, communication overlap
• Develop tooling for experiment tracking, checkpoint management, and reproducibility across long training runs
• Build internal infrastructure that makes researchers productive: fast iteration loops, interpretable training diagnostics, easy ablation tooling
• Collaborate with research to support novel training objectives, causal pre-training pipelines, and online learning loops

Background We're Looking For

• Strong experience with distributed training frameworks: PyTorch Distributed, DeepSpeed, Megatron-LM, or JAX/XLA or equivalent
• Experience making multi-node GPU training reliable in practice, including failures, OOMs, checkpointing, stragglers, and dataloader bottlenecks
• Deep familiarity with systems-level performance: NCCL, InfiniBand/NVLink, mixed precision, memory profiling, communication overlap, and GPU utilization
• Solid engineering fundamentals: reliable systems, clean APIs, good oncall practices
• Comfortable reading and contributing to research code

Nice to Have

• Experience with multi-modal or video model training at scale
• Familiarity with online RL training loops or environment-model co-training
• Experience with Kubernetes, Slurm, or cloud-native GPU orchestration
• Experience with CUDA, Triton, custom kernels, or low-level performance optimization

What You'll Work On

• Build and maintain scalable and high-fidelity physics simulation environments for robot manipulation, contact-rich tasks, and long-horizon physical scenarios
• Design procedural generation pipelines for diverse objects, scenes, task structures, and physical parameters
• Develop environment APIs that support structured interventions: programmatic action injection, counterfactual branching, causal perturbation experiments
• Build real-to-sim and sim-to-real workflows: system identification, domain randomization, sensor modeling, and parameter calibration
• Create simulation benchmarks and evaluation harnesses for policy learning, world models, synthetic data generation, and failure analysis
• Collaborate with robotics engineers to align simulation closely with physical hardware behavior

Background We're Looking For

• Experience building simulation environments for robotics or physical AI: Isaac Gym, MuJoCo, PyBullet, Genesis, or similar
• Strong software engineering skills: clean architecture, performance-aware Python/C++, good API design
• Familiarity with physics engines, contact simulation, rigid and deformable body dynamics
• Experience with procedural content generation or large-scale synthetic data pipelines

Nice to Have

• Experience with GPU-accelerated simulation (Isaac Lab, Isaac Gym, or equivalent)
• Familiarity with structured causal experiments or interventional dataset design
• Experience with photorealistic rendering pipelines (Blender, Omniverse, or similar)

What You'll Work On

• Develop model compression and quantization pipelines for deploying large causal world models on onboard compute (edge GPUs, ARM processors, custom accelerators)
• Build real-time inference engines with strict latency budgets: perception-to-action loops in the tens of milliseconds
• Design and implement streaming inference architectures that process continuous sensor inputs (video, depth, force, proprioception) with low-latency causal state updates
• Optimize memory footprint and compute efficiency for on-robot deployment constraints
• Build testing and validation frameworks for on-robot inference: latency profiling, accuracy degradation analysis, failure mode detection
• Collaborate with robotics engineers on hardware bring-up and integration with real-time control stacks

Background We're Looking For

• Strong experience with model compression: quantization (INT8/INT4), pruning, distillation, and efficient architecture design
• Experience deploying ML models on embedded or edge hardware: Jetson, custom SoCs, FPGAs, or similar
• Deep familiarity with inference optimization frameworks: TensorRT, ONNX Runtime, TFLite, or equivalent
• Understanding of real-time systems, latency constraints, and deterministic execution
• Strong C++ and Python engineering skills

Nice to Have

• Experience with robotic middleware: ROS2, LCM, or similar
• Familiarity with multi-modal model inference: vision, depth, and proprioception jointly
• Experience with hardware-software co-design for inference

What You'll Work On

• Develop manipulation systems for contact-rich tasks: grasping, in-hand manipulation, tool use, assembly, and deformable object handling
• Build and tune low-level controllers: impedance control, force/torque control, compliant motion, and reactive behaviors
• Develop whole-body control frameworks for coordinated loco-manipulation and multi-limb tasks
• Design data collection pipelines for causal intervention experiments: structured task execution, perturbation injection, and outcome logging
• Implement and tune motion planning stacks: trajectory optimization, kinematic planning, real-time reactive planning under uncertainty
• Collaborate with the research team to close the loop between physical experiments and causal model updates

Background We're Looking For

• Strong background in robot manipulation: end-to-end project experience from hardware to high-level task execution
• Deep familiarity with control theory: impedance control, admittance control, force control, optimal control
• Experience with motion planning frameworks: MoveIt, Drake, Pinocchio, or equivalent
• Proficiency in real-time C++ and robotics middleware (ROS2 or equivalent)
• Hands-on experience working with physical robot hardware and debugging hardware-software integration issues

Nice to Have

• Experience with whole-body controllers for humanoid or quadruped robots
• Familiarity with learning-based control: policy learning for manipulation, or RL-based controller tuning
• Experience with tactile sensing, force/torque sensing, or proprioceptive feedback for contact-rich tasks

What You'll Work On

• Lead hardware bring-up for new robot platforms: from component sourcing and assembly through first motion and full system validation
• Design and build mechatronic systems for manipulation research: custom end-effectors, force/torque sensing arrays, tactile sensor integration, instrumented fixtures
• Develop and maintain electrical and power systems: motor drivers, power distribution, safety interlocks, and embedded sensing
• Build reliable hardware-software integration: real-time embedded firmware, driver development, and hardware abstraction layers
• Instrument robot systems for high-quality causal data collection: synchronized multi-camera rigs, force/torque logging, proprioception capture, and environmental sensing
• Maintain and operate robot fleet: preventive maintenance, failure diagnosis, repair, and continuous improvement of hardware reliability

Background We're Looking For

• Strong mechatronics background: mechanical design, electronics, and embedded systems — ideally all three
• Experience with robot hardware bring-up: working with off-the-shelf manipulators, mobile bases, or custom hardware platforms
• Proficiency with CAD tools (SolidWorks, Fusion 360, or equivalent) and experience taking designs through fabrication
• Experience with embedded systems: microcontrollers, real-time firmware, motor control, and hardware communication protocols (EtherCAT, CAN, SPI, I2C)
• Hands-on skills: soldering, wiring, fabrication, and the practical ability to debug hardware failures under time pressure

Nice to Have

• Experience with force/torque sensor integration, tactile sensing, or custom end-effector design
• Familiarity with ROS2 or real-time robotics middleware at the driver/HAL level
• Experience designing instrumented experimental rigs for data collection in research settings