FastStair: Learning to Run Up Stairs with Humanoid Robots

TL;DR

FastStair addresses the challenge of agile yet stable humanoid stair climbing by integrating a GPU-parallel DCM-based foothold planner with a multi-stage reinforcement learning framework. A parallel foothold optimization guides safe contact selection during RL, and post-training splits the policy into high- and low-speed experts that are merged with LoRA to cover the full commanded speed range. Real-world experiments on the Oli robot demonstrate stable stair ascent up to $1.65$ m/s and successful traversal of a 33-step spiral in 12 s, validating both safety and speed. The work highlights how model-based priors can be tightly coupled with learning to achieve robust, fast locomotion on challenging stair terrains.

Abstract

Running up stairs is effortless for humans but remains extremely challenging for humanoid robots due to the simultaneous requirements of high agility and strict stability. Model-free reinforcement learning (RL) can generate dynamic locomotion, yet implicit stability rewards and heavy reliance on task-specific reward shaping tend to result in unsafe behaviors, especially on stairs; conversely, model-based foothold planners encode contact feasibility and stability structure, but enforcing their hard constraints often induces conservative motion that limits speed. We present FastStair, a planner-guided, multi-stage learning framework that reconciles these complementary strengths to achieve fast and stable stair ascent. FastStair integrates a parallel model-based foothold planner into the RL training loop to bias exploration toward dynamically feasible contacts and to pretrain a safety-focused base policy. To mitigate planner-induced conservatism and the discrepancy between low- and high-speed action distributions, the base policy was fine-tuned into speed-specialized experts and then integrated via Low-Rank Adaptation (LoRA) to enable smooth operation across the full commanded-speed range. We deploy the resulting controller on the Oli humanoid robot, achieving stable stair ascent at commanded speeds up to 1.65 m/s and traversing a 33-step spiral staircase (17 cm rise per step) in 12 s, demonstrating robust high-speed performance on long staircases. Notably, the proposed approach served as the champion solution in the Canton Tower Robot Run Up Competition.

Figures (10)

• Figure 1: Full-scale humanoid robot Oli performing agile stair-climbing locomotion: (a) parallel foothold planning in IsaacLab, (b) climbing a staircase with 20 cm step height, (c) Canton Tower Run-up, and (d) high-speed ascent.
• Figure 2: Overview of the FastStair framework. During the pre-training stage, a parallel DCM-based foothold optimizer generates dynamically feasible contacts to guide policy learning via a foothold-tracking reward. In the post-training stage, the pre-trained base model is fine-tuned into high- and low-speed expert policies by expanding the commanded velocity range and adjusting reward weights to mitigate planner-induced conservatism; this decomposition is motivated by the distinct action distributions at high versus low speeds. However, direct switching between experts can cause control discontinuities, so in the LoRA fine-tuning stage we merge their parameters into a single network and fine-tune it with LoRA. The resulting unified policy ensures robust tracking across the full velocity range for deployment on the physical robot.
• Figure 3: Inverted-pendulum stair-climbing process.
• Figure 4: Terrain-perception information. Blue markers show terrain scandots, green the averaged gradient map, light blue the local averaged gradient map, and red the optimal footholds.
• Figure 5: Robot Platform Limx Oli.