Towards Low-gravity Planetary Exploration using RL for Walking, Jumping & In-flight Attitude Control

Added: 2026-05-27

[00:00:00]This video presents our work towards lowgravity planetary exploration using reinforcement learning for walking, jumping, and in-flight attitude control.

[00:00:09]We present a series of reinforcement learning policies that when combined in a taskbased deployment in a simulated Mars environment, enables the robot to overcome obstacles larger than itself terrain where traditional legged robots and rovers would struggle. The system uses walking for normal traversal, forward jumps to clear obstacles, vertical jumps for elevated observation, and attitude control during all flight faces to achieve target orientation for safe controlled landings. In this simulated mission, a waypointbased state machine coordinates policy execution.

[00:00:40]The robot walks or jumps to a specified waypoint while the state machine automatically transitions between policies based on sensor feedback and waypoint commands.

[00:00:51]The quarterbird robot Olympus was designed and optimized for low-gravity jumping and attitude control. The illustrated control architecture was deployed in simulation and on the physical robot with policies generating actions based on task specific observations. The vertical jumping policy is trained to achieve a commanded height. We also investigated the effect of springs on achievable jump height by training a policy with simulated parallel springs. The forward jumping policy is tasked to perform forward jumps to a target position. This task was also investigated with and without springs to understand their effect on jump distance and control authority. The attack control policy is trained to regulate the body orientation by using its legs as reaction masses during flight. Its objective is to maintain the target orientation between takeoff and landing, ensuring safe feet first touchdowns. This was validated through single axis step responses and composite 3D reorientation maneuvers in simulation.

[00:01:48]To showcase the systems capabilities, we commended aggressive inflight orientations that challenge the jumping policy during landing. Although the signal policy typically ensures safe landings, these tests deliberately created difficult conditions to demonstrate both the orientation authority and landing robustness.

[00:02:06]To validate simp transfer of the control policy, we deployed it on the physical robot at ESS orbital robotics lab. The robot was mounted on a rotating rod attached to a air bearing platform floating on an ultra flat floor enabling near frictionless rotation and translation. The policy was evaluated through single axis tests for roll pitch and yaw commanding step changes of 90 and 180° to the target orientation.

[00:03:00]Mounted on the free floating platform, the robot bounces between walls, receiving 180° reorientation command at each push off to maintain feet forward orientation towards the approaching wall. This validates the policy's ability to perform rapid attitude corrections during free flight under dynamic conditions. This work demonstrates that reinforcement learning can enable complex locomotion for planetary exploration in reduced gravity environments. By combining walking, jumping, and attitude control policies, we achieve capabilities that extend beyond those of traditional leg robots and wheeled rovers, opening new possibilities for exploring challenging terrains such as Martian lava tubes.

[00:03:39]Thank you.