The Role

Self-Traversal is the locomotion problem of moving across any surface, in any gravity, of any geometric complexity, with no prior assumption about what the robot will find beneath its feet. It is the foundation capability beneath every hardware embodiment GRAM will ever ship.

Insects already solve it. A stick insect crosses a twig at any angle. A fly walks the underside of a ceiling. An ant traverses bark indistinguishable from a vertical 3D lattice. They operate without elevation maps, without flat-ground priors, without a body-frame assumption that up is fixed. Six legs, redundant contacts, local control, perception conditioned on what is actually beneath their feet. Biology is the existence proof, and four hundred million years of evolution is the prior art.

GRAM's robot is the engineered counterpart. The shape of the technical answer follows from the biology. A multi-legged platform, because reconfiguration on arbitrary structure demands redundant grasp. An RL-trained policy, because the contact schedule across surfaces of unknown geometry is too combinatorial for hand-authored gaits or trajectory optimization alone. Vision-coupled control, because the next foothold has to come from raw perception of what is actually there, not from a height map that assumes flat ground exists. Gravity-agnostic, because up is whatever direction the body happens to be pointing.

You will own the policy that makes Self-Traversal real, in simulation, on hardware, and across the gap between them.

Key Responsibilities

• Own the Self-Traversal locomotion policy end to end. Train in simulation, deploy on hardware, close the sim-to-real gap on a contact-rich, non-planar platform.
• Design contact-aware RL training environments and curricula for arbitrary 3D structure, with domain randomization across surface geometries, contact mechanics, and gravitational orientations.
• Architect the vision-coupled footstep-selection stack so next-foothold decisions are conditioned on raw perception of arbitrary geometry rather than precomputed elevation maps.
• Co-design with mechatronics and adhesion teams so the controller exploits the gripper, microspine, or compliant-foot mechanism.
• Extend Self-Traversal to multi-robot configurations: several robots co-occupying a single structure, deconflicting overlapping coverage in real time without central planning.

About You

You can create.

Basic Qualifications

• Demonstrated work in robot locomotion: research output, hardware deployment, open-source contribution, or production system. MS in Robotics, CS, Mechanical, or Electrical Engineering preferred for production candidates and waivable for strong artifact evidence.
• Evidence of a learned policy you have personally taken from simulation onto physical hardware, at any scale.
• Working fluency in Python and at least one modern legged-robot stack: Isaac Lab, legged_gym, rsl_rl, MuJoCo (MJX or MuJoCo MPC), Drake, Pinocchio, OCS2, or Crocoddyl. C++ proficiency expected for production candidates and welcome to develop for early-career applicants.
• Foundational understanding of modern RL (PPO, SAC, off-policy methods) and classical contact mechanics. Pure-simulation RL with no hardware deployment is disqualifying. Pure-MPC backgrounds with no exposure to learned policies are disqualifying.

Preferred Qualifications

• PhD focus on legged locomotion, learned control, or contact-rich robotics, with publications at RSS, CoRL, ICRA, IROS, NeurIPS, ICML, or ICLR.
• Direct lineage from one or more of: ETH RSL, MIT Improbable AI, Berkeley Hybrid Robotics, Stanford IPRL, NVIDIA Isaac, Oxford ORI, CMU LeCAR, UCSD (Wang lab), JPL LEMUR / Parness, Cutkosky / BDML at Stanford, Ramezani at Northeastern, Penn Kodlab, KAIST CLS.
• Vision-conditioned locomotion experience: footstep selection conditioned on RGB, depth, or event cameras rather than elevation maps.
• Non-planar contact experience: climbing, inversion, microspine or gecko-adhesion grippers, asteroid-surface mobility, or any setting where the gravity vector relative to the body is not constant.
• Familiarity with whole-body contact-rich analytical control (TSID, Pinocchio, OCS2, Crocoddyl). The stack may layer analytical contact-force regulation under the learned policy at the gripper interface.