KiRAS: Keyframe Guided Self-Imitation for Robust and Adaptive Skill Learning in Quadruped Robots

Xiaoyi Wei; Peng Zhai; Jiaxin Tu; Yueqi Zhang; Yuqi Li; Zonghao Zhang; Hu Zhou; Lihua Zhang

cs.ROMar 16, 2026

KiRAS: Keyframe Guided Self-Imitation for Robust and Adaptive Skill Learning in Quadruped Robots

Authors: Xiaoyi Wei, Peng Zhai, Jiaxin Tu, Yueqi Zhang, Yuqi Li, Zonghao Zhang, Hu Zhou, Lihua Zhang

Organizations: College of Intelligent Robotics and Advanced Manufacturing, Fudan University, China. · Power China Huadong Engineering Corporation Limited, China.

Abstract

With advances in reinforcement learning and imitation learning, quadruped robots can acquire diverse skills within a single policy by imitating multiple skill-specific datasets. However, the lack of datasets on complex terrains limits the ability of such multi-skill policies to generalize effectively in unstructured environments. Inspired by animation, we adopt keyframes as minimal and universal skill representations, relaxing dataset constraints and enabling the integration of terrain adaptability with skill diversity. We propose Keyframe Guided Self-Imitation for Robust and Adaptive Skill Learning (KiRAS), an end-to-end framework for acquiring and transitioning between diverse skill primitives on complex terrains. KiRAS first learns diverse skills on flat terrain through keyframe-guided self-imitation, eliminating the need for expert datasets; then continues training the same policy network on rough terrains to enhance robustness. To eliminate catastrophic forgetting, a proficiency-based Skill Initialization Technique is introduced. Experiments on Solo-8 and Unitree Go1 robots show that KiRAS enables robust skill acquisition and smooth transitions across challenging terrains. This framework demonstrates its potential as a lightweight platform for multi-skill generation and dataset collection. It further enables flexible skill transitions that enhance locomotion on challenging terrains.

Figures & tables

Fig. 1: Deployment of KiRAS on Solo-8 and Unitree Go1 robots across diverse environments. (a) Solo-8 robot traverses various obstacles by flexibly switching learned skills. (b) Solo-8 robot exploits crawl for concealment and stilt for expanded field of view. (c) In outdoor unstructured environments, Solo-8 robot demonstrates strong robustness. (d) Unitree Go1 robot acquires bipedal skills to climb a 10 cm step.

Fig. 2: Training pipeline. KiRAS consists of an end-to-end adaptive multi-skill learning framework and a skill initialization module. The former includes a keyframe-guided skill learning module (used only in skill learning stage), a proprioception estimator, and a PPO actor-critic architecture. The policy outputs joint position targets, which are converted to torques via a PD controller. The Skill Initialization Technique is employed at environment reset to prevent overfitting to simpler skills. The purple box illustrates different skills being trained on different terrains. During deployment, only the networks in the yellow boxes are used, and skill switching is controlled via joystick input.

Fig. 3: Details of Premium Trajectory Selector.

Skills	Walk	Crawl	Stilt	Pitch-up	Pitch-down
Base Height (cm)	20	10	30	20	20
Pitch Angle (deg)	0	0	0	-15	15

TABLE I: Parameters of Different Skills.

Fig. 4: Heatmap of ablation and comparison experiment results. The vertical axis ”C” denotes the cosine similarity between each skill trajectory and its corresponding keyframe, with higher values indicating greater similarity. ”S” and ”D” represent the success rates of traversing steps and discrete footholds, respectively, where higher values indicate more robust policies.

Fig. 5: Comparative results. (a) KiRAS w/o Skill Initialization learns 5 skills (walk, crawl, stilt, pitch-up, pitch-down) with little distinction. (b) DreamWaQ’s crawl and pitch-down use knee-ground contact (red circles). (c) KiRAS w/o MuC fails to lift legs over obstacles.

Fig. 6: Real-world experiments validating skill acquisition with KiRAS on the Solo-8 (a) and Unitree Go1 (b) robots. Each black dashed box shows skill execution on hardware, with the left image showing the real robot and the right showing the keyframe used for training. In all experiments, robots are oriented to the right as the forward direction.

Fig. 7: Sampling probability evolution for each skill during training, computed using 4 random seeds. Curves and shaded areas represent the mean and standard deviation across seeds, respectively.

Fig. 8: Solo-8 robot traversing diverse and challenging terrains in indoor and outdoor environments with different skills.

Skills	Walk	Crawl	Stilt	Pitch-up	Pitch-down
Slope	0.84	0.68	0.80	0.70	0.9
Upstairs	0.80	0.54	0.44	0.88	0.58
Downstairs	0.88	0.46	0.70	0.94	0.56

TABLE II: Success rates of Solo-8 robot traversing terrains with different skills. The highest rates are in bold.

Fig. 9: (a) Transition from the walk skill to the biped skill. (b) Keyframe for the biped skill used for imitation. (c) Snapshots of executing the biped skill to traverse various indoor and outdoor terrains.

Observation Term	$\mathbf{o_{t}^{\mathrm{prop}}}$	$\mathbf{o_{t}^{V}}$	$\mathbf{\Phi(o_{t})}$
Base linear velocity		✓
Base angular velocity		✓
Foot height		✓
Projected gravity		✓
Base orientation	✓	✓	✓
Command	✓	✓

TABLE III: Observation components.

Reward name	Coefficient
Termination penalty $r_{t}^{T}$
collision	-0.5
Regularization reward $r_{t}^{R}$
feet_contact_forces	-1.0
action_rate	-0.2
torques	$-2.5\times 10^{-5}$

TABLE IV: Reward terms and coefficients.

Parameter	Range / Value
Base payload perturbation	$[-0.5,\,0.25]$ kg
Base CoM shift (per axis)	$[-0.02,\,0.05]$ m
Contact friction coefficient	$[0.5,\,1.25]$
External push velocity (base)	$[-0.1,\,0.1]$ m/s
Push interval	$5$ s
Actuation delay	$[0,\,15]$ ms

TABLE V: Domain randomization settings.

Network	In/Out	Hidden
Actor	$42\rightarrow 8$	[128,128,128]
Task critic $V_{\phi_{1}}$	$107\rightarrow 1$	[128,128,128]
Optimization critic $V_{\phi_{2}}$	$107\rightarrow 1$	[128,128,128]
ECE encoder	$124\rightarrow 19$	[128,64]
ECE decoder	$13\rightarrow 31$	[64,128]
Imitation discriminator	$38\rightarrow 1$	[512,256]

TABLE VI: Network architectures for training Solo-8 policy.

Skills	Walk	Stilt	Crawl	Pitch-up	Pitch-down
DTW $\downarrow$
Keyframe	15.67	15.19	20.11	18.03	19.04
Dataset	16.42	15.84	24.18	19.53	19.96
Cosine Similarity $\uparrow$
Keyframe	0.978	0.974	0.987	0.965	0.976
Dataset	0.963	0.942	0.966	0.949	0.953

TABLE VII: Comparison of DTW (lower is better) and Cosine Similarity (higher is better) across skills.

Robot Skill Learning

Robot skill learning studies how robots acquire, adapt, and combine reusable behaviors from demonstrations, interaction, and other data.

See all 320 papers

Agile perceptive multi-skill locomotion for quadrupedal robots in the wild
Jul 15, 2026Jun-Gill Kang, Jaehyun Park, Tae-Gyu Song +3Terrain-Aware Robot Locomotion Robot Locomotion
1Agency for Defense Development, Daejeon 34186, Republic of Korea · Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea · 3DIDEN Robotics, Seoul 04799, Republic of Korea +2
Teaching a Robot Dog New Tricks: Diverse Quadruped Skills via Combined Reinforcement and Imitation Learning with Adversarial Task Selection
Oct 6, 2026Lemon Foxmere, Anthony Furman, Yizheng Du +3Robot Skill Learning Multi-Task RL
HARE Lab, University of California at Santa Cruz, Santa Cruz, CA 95064 USA · AIEA Lab, University of California at Santa Cruz, Santa Cruz, CA 95064 USA
LatentMimic: Terrain-Adaptive Locomotion via Latent Space Imitation
Apr 7, 2026Zhiquan Wang, Yunyu Liu, Dipam Patel +3Terrain-Aware Robot Locomotion Quadruped Locomotion
Department of Computer Science, Purdue University, USA

Jul 15, 2026J/K move · Enter open · S save

Agile perceptive multi-skill locomotion for quadrupedal robots in the wild

Jun-Gill Kang, Jaehyun Park, Tae-Gyu Song, Joon-Ha Kim+2

1Agency for Defense Development, Daejeon 34186, Republic of Korea · Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea · 3DIDEN Robotics, Seoul 04799, Republic of Korea+2

Terrain-Aware Robot Locomotion Robot Locomotion RL for Legged Locomotion Quadruped Locomotion

Open paper AskPDF arXiv

Enabling quadrupedal robots to traverse complex terrains-from rugged outdoor environments to urban landscapes-requires seamless integration of multiple motor skills, smooth transitions between gaits, and high-speed perceptive locomotion using only onboard sensors. We present APT-RL (Action Pretrained Transformer-based Reinforcement Learning), a unified framework that enables multi-skill locomotion to achieve high-speed traversal in complex environments through autonomous skill transitions utilizing only onboard perception and computation. Our approach generates large-scale, feature-rich 2D motion datasets through trajectory optimization with simplified dynamics. These datasets enable training of diverse, reusable locomotion skills that transfer effectively to a real quadruped robot operating on complex uneven terrains. The resulting high-quality skills serve as strong priors for efficient learning of complex downstream tasks and extend naturally to 3D environments, enabling smooth, high-speed multi-skill locomotion in deployed policy. Real-world experiments demonstrate the framework's capabilities: the robot performs agile maneuvers through complex indoor obstacles and outdoor wild environments, including dynamic drop-down maneuvers that reach instantaneous peak speeds of up to 6 meters per second. A single onboard policy enables robust traversal of diverse obstacles, including stairs, hurdles, stepping stones, gaps, and fallen branches, demonstrating the versatility and effectiveness of our approach.

Abstract

Robot Skill Learning

Explore similar work

Agile perceptive multi-skill locomotion for quadrupedal robots in the wild

Teaching a Robot Dog New Tricks: Diverse Quadruped Skills via Combined Reinforcement and Imitation Learning with Adversarial Task Selection

LatentMimic: Terrain-Adaptive Locomotion via Latent Space Imitation