🤖 AI

Diffusion-Based Generation of Gait Trajectories

This paper proposes and evaluates conditional diffusion models, specifically a controllable diffusion transformer with adaptive normalization and classifier-free guidance, to generate realistic and personalized lower-limb gait trajectories conditioned on parameters like step length, offering a scalable alternative to traditional hand-crafted methods for wearable robotics and rehabilitation.

Original authors: Damian Benasco, Juan Carballeira-Lopez, Jaime Ramos-Rojas, Julio S. Lora-Millan, Antonio J. Del-Ama, David Rodriguez-Cianca, Pablo Lanillos

Published 2026-09-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Damian Benasco, Juan Carballeira-Lopez, Jaime Ramos-Rojas, Julio S. Lora-Millan, Antonio J. Del-Ama, David Rodriguez-Cianca, Pablo Lanillos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Walking is a feat of complex coordination that the human body performs without conscious thought, yet for people with mobility impairments, this natural rhythm can be lost or distorted. In the field of rehabilitation robotics, scientists are working to build wearable machines, such as exoskeletons, that can help restore this lost movement. For these machines to work effectively, they need a map of how a person should move. This map, known as a reference trajectory, must be perfectly tailored to the individual's body shape, their specific injury, and their therapeutic goals. If the machine moves in a way that feels unnatural or ignores the user's unique physical limits, it can be uncomfortable or even harmful. The challenge lies in creating these movement maps on the fly, adapting them instantly to different people and different walking conditions, rather than relying on rigid, pre-programmed templates that fail to capture the nuance of human motion.

To solve this, a team of researchers from Spain has turned to a type of artificial intelligence known as a diffusion model. Imagine a process where an image is slowly covered in static noise until it is unrecognizable, and then a computer learns to reverse that process, removing the noise step by step to reveal a clear picture. In this study, the researchers applied this same logic to human movement. Instead of pixels, they used data points representing the angles of the joints in the legs. They trained the computer on thousands of recorded walking cycles from healthy volunteers, teaching it to understand the underlying patterns of a natural stride. The goal was to see if the AI could generate new, realistic walking paths that were not just copies of the data it had seen, but were instead customized to specific instructions, such as taking a shorter or longer step.

The researchers tested two different approaches using data from 22 healthy adults who walked at two distinct paces: a short step and a medium step. The first approach was a baseline model that learned the general shape of a walk but did not have a way to be told exactly how long the steps should be. As expected, this model produced walking patterns that looked real, but it tended to settle on an average step length, effectively ignoring the specific request to walk with a shorter or longer stride. It was like a musician who could play a song beautifully but could not change the tempo when asked.

The second approach, which the team developed as a more advanced version, introduced a mechanism to control the output directly. By feeding the AI specific physical parameters, such as the desired step length, the model learned to adjust the entire movement pattern to match that instruction. When tested, this controllable model successfully generated walking trajectories that matched the requested step lengths. For the medium-length steps, the generated movements were highly accurate, closely matching the real human data in terms of joint angles and symmetry between the left and right legs. The model produced a step length of roughly 68.6 centimeters, which aligned well with the target.

However, the results were not perfect for every condition. When the researchers asked the model to generate very short steps, the accuracy dropped. The AI produced steps that were shorter than the target, but the movements were less precise and showed more variation from person to person. The researchers found that this was because short steps are naturally more inconsistent across different people; some people shuffle, while others take small, quick steps. Because the data for short steps was so varied, the AI struggled to find a single, clear pattern to follow and tended to revert toward a middle ground. This resulted in a slight underestimation of how much the joints should move, particularly in the knee and hip. Despite this limitation, the study demonstrated that the advanced model could separate the two types of walking clearly, whereas the simpler model could not distinguish between them at all.

The findings suggest that while artificial intelligence can now generate highly realistic and customizable walking paths, the task becomes significantly harder when the requested movement is less common or more variable across individuals. The researchers concluded that their new method holds promise for creating personalized guides for robotic exoskeletons, allowing these machines to adapt to a patient's specific needs in real time. They noted that for the technology to be ready for clinical use, future work will need to address how to control other factors, such as walking speed and individual body dimensions, and how to handle the full range of human walking styles. For now, the study proves that diffusion models can move beyond simple imitation to become tools that can generate new, plausible ways for humans to walk, tailored to the unique geometry of each person's body.

Technical Summary: Diffusion-Based Generation of Gait Trajectories

Problem Statement
The generation of musculoskeletal gait trajectories conditioned on patient-specific parameters remains a critical challenge for wearable robotics and rehabilitation. Assistive systems, particularly lower-limb exoskeletons, require reference trajectories that adapt to individual morphology and therapeutic goals while preserving biomechanical realism. Traditional approaches relying on hand-crafted gait templates or optimization procedures often scale poorly across different subjects and walking conditions. While recent AI-driven methods like reinforcement learning have shown promise in adaptive control, and generative models like diffusion models have succeeded in motion synthesis, their specific application to clinical gait generation conditioned on physical parameters remains largely unexplored.

Methodology
The authors frame gait synthesis as a denoising diffusion process, proposing two distinct model architectures trained on a dataset of 4,590 gait cycles from 22 healthy subjects. The dataset includes 21 lower-limb joint angles and biomechanical keypoints derived from VICON motion capture, categorized into two distinct step-length classes: short (< 48 cm) and medium (> 52 cm).

Baseline Model: A keypoint-conditioned Transformer diffusion model (following the MDM architecture) treats each frame as a token. It conditions on biomechanical keypoint signals (gait-phase pulses, event markers) via channel concatenation and predicts clean samples over 500 timesteps.
Controllable Diffusion Transformer (DiT): To enable conditioning on physical gait parameters (specifically step length), the authors introduce a variant incorporating AdaLN-Zero. In this architecture, a vector composed of timestep and gait-parameter embeddings produces per-layer shift, scale, and gate modulations, replacing standard layer normalization. Physical parameters are injected at every layer, while kinematic pulse conditioning remains at the input.
Training and Guidance: The DiT utilizes classifier-free guidance (CFG) via a learnable null embedding (substituted during training with probability $p=0.25$ ) to control the generation. The training objective minimizes a composite loss function ( $L$ ) combining reconstruction ( $L_{mse}$ ), periodicity ( $L_{cyc}$ ), smoothness ( $L_{vel}$ ), and a forward-kinematics foot-trajectory term ( $L_{fk}$ ), with weights ramped via curriculum learning.

Key Results
The study evaluates the models using Pearson Correlation Coefficient ( $R$ ), Normalized RMSE (NRMSE), Range of Motion (ROM) error, and Symmetry Index (SI).

Baseline Performance: The baseline model captures general gait dynamics ( $R=0.777$ ) but converges to the dataset mean step length (57.4 cm), lacking specific controllability.
DiT Performance (Medium Steps): The DiT achieves high fidelity for medium steps ( $R=0.838$ , NRMSE=0.920) with a generated step length of 68.6 cm. ROM errors are within $\pm6^\circ$ , and symmetry is well-preserved.
DiT Performance (Short Steps): Conditioning for short steps successfully shifts kinematics to the target mean (36.4 cm, $R=0.813$ ). However, this class exhibits higher error (NRMSE=1.225) and systematic underestimation of ROM (approx. $10^\circ$ for knee, $6^\circ$ for hip). The authors attribute this to greater inter-subject variability in short strides (knee ROM CoV $\approx 42\%$ vs. 26%) and the bias introduced by the CFG null-embedding, causing the MSE-trained diffusion to regress toward the mean.
Symmetry: Symmetry is preserved in the medium-step class but degrades in the short-step class (hip/knee SI $\approx 16–17\%$ ).

Significance and Claims
The paper claims that diffusion-based models can generate realistic, periodic lower-limb gait trajectories while enabling controllability over specific gait characteristics, such as step length. The proposed DiT architecture demonstrates the potential for personalized gait synthesis in assistive robotics, offering a method to generate kinematic references for adaptive exoskeleton controllers.

However, the authors maintain a modest stance regarding clinical deployment. They note that step length alone is insufficient to distinguish all clinically relevant gait patterns (e.g., a slow shuffle vs. a fast short stride may share step lengths but differ in cadence and speed). Consequently, the work is presented as a proof of concept for diffusion-based gait generation, with future work identified as necessary to address conditioning on speed, subject-specific body dimensions, and continuous multi-parameter conditioning.

Technical Summary: Diffusion-Based Generation of Gait Trajectories

More like this