Minerva Talent Program

Develop and refine locomotion policies for a next-generation humanoid robot platform, enabling robust walking, stair climbing, and terrain adaptation in unstructured industrial environments. You will work at the intersection of sim-to-real reinforcement learning and real-world deployment on physical hardware.

RESPONSIBILITIES

•  Train locomotion policies in simulation (Isaac Sim / MuJoCo) using reinforcement learning, with a focus on robustness to terrain variation, payload changes, and external disturbances

•  Design and implement sim-to-real transfer pipelines, including domain randomization and system identification, to deploy policies on the physical humanoid platform

•  Develop and benchmark locomotion primitives (flat walking, incline traversal, step climbing) against performance targets for speed, stability, and energy efficiency

•  Instrument hardware tests and build evaluation frameworks to quantify policy performance on the physical robot

•  Collaborate with the dynamic motion and perception teams to integrate locomotion with higher-level task planners and agile maneuver capabilities

QUALIFICATIONS

•  Graduate student (MSc or PhD) in Robotics, CS, or related field with coursework in reinforcement learning and robot control

•  Strong Python proficiency; experience with PyTorch and at least one robotics simulator (Isaac Sim, MuJoCo, PyBullet)

•  Familiarity with legged locomotion literature (e.g., policies from Hybrid RL, AMP, or similar frameworks)

•  Experience deploying learned policies on physical hardware is a strong plus

•  Comfortable working in a fast-paced startup environment with ambiguous problem definitions

Push the boundaries of humanoid agility by developing learning-based controllers for highly dynamic motions: running, jumping, climbing over obstacles, and coordinated locomanipulation. You will train policies that enable a humanoid robot to perform parkour-inspired maneuvers and interact with its environment at speed, targeting deployment in industrial settings where navigating cluttered, multi-level structures is critical.

RESPONSIBILITIES

•  Train end-to-end reinforcement learning policies for dynamic humanoid behaviors including running, jumping, vaulting, and climbing over obstacles in simulation

•  Develop locomanipulation skills that coordinate whole-body motion with arm interactions (e.g., grabbing rails while climbing, bracing against surfaces, opening hatches while balancing)

•  Design reward functions and curriculum learning strategies that progressively build from basic dynamic gaits to complex parkour-style motion sequences

•  Implement sim-to-real transfer techniques (domain randomization, dynamics augmentation) to bridge the gap for high-impact dynamic motions on physical hardware

•  Build evaluation benchmarks for agile locomotion covering success rate, robustness to perturbation, and generalization across obstacle configurations

QUALIFICATIONS

•  Graduate student (MSc or PhD) in Robotics, CS, or related field with strong experience in reinforcement learning for locomotion or manipulation

•  Deep familiarity with physics simulators (Isaac Sim, MuJoCo) and training frameworks for contact-rich, dynamic tasks

•  Experience with whole-body control, motion imitation learning (e.g., AMP, DeepMimic), or agile locomotion research

•  Strong Python and PyTorch proficiency; C++ experience for real-time deployment is a plus

•  Comfort with high-risk hardware experiments and iterating rapidly between simulation and physical testing

Build and optimize a multi-camera visual SLAM pipeline for a new humanoid robot platform equipped with several onboard cameras. The goal is to enable reliable real-time localization and dense mapping in GPS-denied industrial facilities such as refineries, offshore platforms, and processing plants.

RESPONSIBILITIES

•  Develop and integrate a multi-camera visual SLAM system that fuses inputs from the robot's camera array for robust 6-DOF pose estimation

•  Implement loop closure, relocalization, and map management strategies tailored to repetitive industrial environments (pipes, corridors, symmetric structures)

•  Optimize the pipeline for real-time performance on the robot's onboard compute, profiling and reducing latency across the perception stack

•  Build evaluation tools and benchmark datasets using data collected on the physical platform in lab and field environments

•  Collaborate with the navigation and locomotion teams to feed accurate localization into path planning and gait adaptation modules

QUALIFICATIONS

•  Graduate student (MSc or PhD) in Computer Vision, Robotics, or related field with strong foundations in multi-view geometry and SLAM

•  Experience with visual or visual-inertial SLAM systems (ORB-SLAM, VINS-Mono, Kimera, or similar)

•  Proficiency in C++ and Python; experience with ROS2 and OpenCV

•  Familiarity with camera calibration, multi-camera extrinsic estimation, and sensor fusion

•  Experience working with real sensor data on physical robot platforms is a strong plus

Develop semantic perception capabilities that enable a humanoid robot to understand and reason about its surroundings in complex industrial environments. This role focuses on turning raw multi-camera imagery into actionable scene representations - traversability maps, hazard detection, and object-level understanding - that feed directly into the robot's autonomous navigation stack.

RESPONSIBILITIES

•  Build and fine-tune semantic segmentation and object detection models for industrial scene understanding (walkable surfaces, obstacles, stairs, hazards, equipment)

•  Develop a multi-camera fusion pipeline that combines per-camera semantic outputs into a unified 3D semantic map around the robot

•  Design traversability estimation modules that classify terrain and predict safe footholds using both geometric and semantic cues

•  Create data collection and annotation workflows for industrial environments, including synthetic data generation from simulation

•  Integrate semantic outputs with the SLAM and path planning systems to enable context-aware autonomous navigation.

QUALIFICATIONS

•  Graduate student (MSc or PhD) in Computer Vision, Machine Learning, or Robotics with experience in semantic segmentation or 3D scene understanding

•  Strong proficiency with PyTorch and modern vision architectures (transformers, foundation models for segmentation)

•  Experience with 3D point cloud processing, depth estimation, or multi-view 3D reconstruction

•  Familiarity with ROS2 and deploying perception models on edge hardware (NVIDIA Jetson or similar)

•  Interest in bridging perception and planning; prior exposure to navigation stacks is a plus