Learning to Jump from Pixels

TL;DR

The paper tackles vision-based agile locomotion over discontinuous terrain by introducing Depth-based Impulse Control (DIC), a hierarchical framework that pairs a vision-driven high-level policy with a model-based low-level impulse controller to produce real-time jumping trajectories. By training with model-free reinforcement learning and leveraging a Raibert-inspired impulse strategy for foot-ground interactions, DIC achieves robust gap-crossing behaviors without relying on dynamics randomization for sim-to-real transfer. The approach yields emergent, adaptable gaits and demonstrates cross-domain performance, with successful real-world gap crossings of up to 26 cm under favorable sensing, though transfer limits remain due to state estimation drift and foot-slip. Overall, DIC advances vision-guided agile locomotion on quadrupeds by effectively integrating perception, learning, and impulse-based control.

Abstract

Today's robotic quadruped systems can robustly walk over a diverse range of rough but continuous terrains, where the terrain elevation varies gradually. Locomotion on discontinuous terrains, such as those with gaps or obstacles, presents a complementary set of challenges. In discontinuous settings, it becomes necessary to plan ahead using visual inputs and to execute agile behaviors beyond robust walking, such as jumps. Such dynamic motion results in significant motion of onboard sensors, which introduces a new set of challenges for real-time visual processing. The requirement for agility and terrain awareness in this setting reinforces the need for robust control. We present Depth-based Impulse Control (DIC), a method for synthesizing highly agile visually-guided locomotion behaviors. DIC affords the flexibility of model-free learning but regularizes behavior through explicit model-based optimization of ground reaction forces. We evaluate the proposed method both in simulation and in the real world.

Figures (15)

• Figure 1: We propose a system architecture called Depth-based Impulse Control (DIC) to enable the Mini Cheetah to jump over wide gaps using depth data captured from an onboard camera.
• Figure 2: Depth-based Impulse Control (DIC; right) maps robot state and vision to a whole-body trajectory in contrast to previous work that directly predicts joint positions (left). The low-level MPC+WBIC controller enables tracking of highly dynamic whole-body trajectories.
• Figure 3: High-level gait prediction network
• Figure 4: (a) Visually guided fixed gait policies significantly outperform blind policies and are close to the "ideal" theoretical limit. Shaded regions indicate standard error of the mean. (b) A comparison of performance among policies trained with fixed gait and unconstrained gait demonstrates the flexibility and dynamic range of our method.
• Figure 5: Contact schedule generated by our variable gait policy. Given a terrain observation (top), the policy modulates body velocity (middle) and contact duration (bottom) to traverse 30cm gaps.