Controls Engineer, Locomotion

The Mission

GRAM is a self-replication company creating machine labor for the physical economy.

Our first research frontier is self-preservation: the base case of physical self-replication. We are building a new class of machines that can survive, coordinate, and recover without humans. We believe scalable machine labor requires more than single-agent task generality or machines shaped in our image.

Our work spans hardware, controls, reinforcement learning, multi-agent coordination, materials science, evaluation, and world models. Join us to solve closure and multi-agent environment generality in industrial domains where demand for labor is effectively unbounded.

It is our mission to make humanity galactic.

Self-Traversal is the locomotion problem at the center of GRAM: moving across arbitrary 3D structure, in any body orientation, with no assumption that the next contact patch is flat, known, or floor-like.

You will own the locomotion stack that makes this real on hardware. The near-term benchmark is simple to state and hard to achieve: a multi-legged robot should cover the usable structure of a complex steel lattice structure from a single placement, using learned policies, local contact intelligence, and perception-conditioned foothold selection.

The technical shape is specific: redundant contact on a multi-legged platform; learned contact schedules that generalize across substrate geometry; vision-conditioned local foothold selection from raw geometry; and gravity-agnostic stability across vertical, lateral, and inverted orientations.

This is not a pure simulation role. You will train policies, deploy them on physical robots, break them against real contact mechanics, and close the loop between simulator, controller, perception, adhesion, and hardware.

• Own GRAM's Self-Traversal locomotion policy from simulation through hardware deployment.
• Build contact-aware RL environments and curricula for arbitrary 3D structure, with domain randomization across geometry, contact mechanics, adhesion, and gravity/body orientation.
• Develop vision-conditioned foothold and path-selection systems that use raw geometry and local perception rather than flat-ground or height-map assumptions.
• Work with mechatronics, firmware, and adhesion teams so the controller exploits the actual foot, gripper, microspine, magnetic, or compliant contact mechanism.
• Create evaluation loops for sim-to-real transfer, coverage, recovery, failure classification, graceful degradation under actuator/sensor/contact failures, and hardware regressions.
• Extend locomotion toward multi-robot traversal, where several robots occupy one structure and coordinate coverage without centralized micromanagement.

• You have built or materially contributed to a robot locomotion stack on real hardware.
• You have personally taken a learned policy, controller, or planning stack from simulation into physical deployment.
• You have worked with multi-legged or contact-rich platforms: hexapods, RHex-like systems, quadrupeds, climbing robots, inspection robots, or hardware that must reason through redundant contacts.
• You are fluent in Python and comfortable in at least one modern robotics stack:
  Isaac Lab, , , Mu Jo Co , Mu Jo Co  MPC, Drake, Pinocchio, OCS2, Crocoddyl, ROS2, or an equivalent internal stack.
• You understand both modern reinforcement learning and classical contact mechanics. You do not need to be doctrinaire about either; we care about what survives contact with hardware.
• You can debug across abstraction layers: policy behavior, contact model, perception artifact, actuator limit, firmware timing, adhesion failure, and mechanical failure.

• Publications, open-source work, or deployed systems in legged locomotion, learned control, contact-rich robotics, climbing robotics, or sim-to-real transfer, especially around RSS, CoRL, ICRA, IROS, NeurIPS, ICML, or ICLR.
• Experience with vision-conditioned locomotion, foothold selection, or perception-in-the-loop control using RGB, depth, event cameras, tactile sensing, or local geometry.
• Work on non-planar contact: climbing, inversion, microspines, gecko-style adhesion, magnetic adhesion, compliant feet/grippers, asteroid mobility, or any system where the gravity vector relative to the body is not fixed.
• Depth in sim-to-real policy generalization across substrate geometries, not only policy performance on a fixed benchmark environment.
• Familiarity with whole-body contact-rich analytical control such as TSID, Pinocchio, OCS2, or Crocoddyl, including hybrid stacks where analytical contact-force regulation sits under a learned policy.
• Relevant lab or project lineage such as ETH RSL, MIT Improbable AI, Berkeley Hybrid Robotics, Stanford IPRL, NVIDIA Isaac, Oxford ORI, CMU LeCAR/Biorobotics, UCSD Wang Lab, JPL LEMUR/Parness, Stanford BDML/Cutkosky, Penn Kodlab, KAIST CLS, or comparable industrial robotics teams.
• Field deployment scars: you have watched the robot fail somewhere inconvenient and…