End-to-End Control of a Powered Knee-Ankle Prosthesis Towards Unified, Tuning-Free Assistance

arXiv:2606.07902 · cs.RO · Submitted 2026-06-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "End-to-End Control of a Powered Knee-Ankle Prosthesis Towards Unified, Tuning-Free Assistance".

Dev: Here's the gist of the paper. Powered prostheses conventionally rely on impedance controllers that require extensive manual tuning and explicit mode classification This work presents real-time deployment of an end-to-end prosthesis controller that estimates continuous actuator signals from onboard sensors, eliminating the need for intent classifiers and subject-specific tuning Pruning redundant reasoning tokens The study demonstrates real-time end-to-end control of a powered knee-ankle prosthesis across five locomotion modes The controller adapted to walking speed, distinguished modes, and generalized stair transitions to unseen limb-leading conditions How the training signal is built A multi-terrain locomotion dataset was compiled from prior experiments involving the OSL under real-time impedance control Data from 18 individuals with TF amputation were collected, spanning level walking (LW), ramp ascent (RA) and ramp descent (RD), and stair ascent (SA) and stair descent (SD) The training protocol used Adam with mean squared error loss and early stopping to train the Temporal Convolutional Network Locomotion modes were resampled each epoch to a fixed distribution, with 60% LW, 10% per remaining mode Where the gains come from The controller reproduced the training-data scaling of peak ankle torque with walking speed during level walking This scaling relationship held for able-bodied participants despite the absence of a bypass adapter in training, suggesting generalization to unseen users Integrated gradients showed the model converged on biomechanically interpretable input dependencies without architectural priors The controller produced mode-distinct joint behaviors across ramp conditions, supporting its ability to detect environmental context shifts Joint Kinematics and Kinetics During Stair Ambulation Transitions The system was evaluated against two hypotheses: that the controller reproduces the assistance scaling patterns present in its training data and that it detects environmental context shifts and generates distinguishable joint behaviors across locomotion modes The transition cue in both stair sequences was the user’s residual thigh lift, captured by the thigh IMU as they prepared to step onto a stair This indicates that the controller learned to interpret user intent directly from onboard mechanical sensors without requiring vision or depth cameras The results represent a promising step towards a unified, tuning-free prosthesis controller that could improve deployment scalability The small sample size limits the generalization of these findings, suggesting the results should be viewed as preliminary This study demonstrated real-time end-to-end control of a powered knee-ankle prosthesis across five locomotion modes The controller adapted to walking speed, distinguished modes, and generalized stair transitions to unseen limb-leading conditions The controller produced mode-distinct joint behaviors across ramp conditions, supporting its ability to detect environmental context shifts The results represent a promising step towards a unified, tuning-free prosthesis controller that could improve deployment scalability The small sample size limits the generalization of these findings, suggesting the results should be viewed as preliminary

Paper summary: Rosa: So wrapping up this discussion on "End-to-End Control of a Powered Knee-Ankle Prosthesis Towards Unified, Tuning-Free Assistance". The paper demonstrates real time end-to-end control across five locomotion modes, adapting to walking speed and distinguished modes.

Dev: It also showed the controller can generalize stair transitions to unseen limb leading conditions and produce mode distinct joint behaviors during ramp conditions.

Taro: Basically, it means they achieved a unified controller that can handle different terrains without needing complex manual tuning or explicit intent classifiers.

Rosa: The implication for the wider field is that this approach could significantly improve how quickly these devices can be deployed because you’re not stuck with subject specific tuning for every single user.

Dev: And the core message is that continuous estimation from onboard sensors can replace the need for those traditional, more labor-intensive control methods.

Taro: It points toward a future where prostheses are much more integrated and adaptive to whatever the user is doing in their immediate environment.

Conclusion: Rosa: So we're looking at this paper, "End-to-End Control of a Powered Knee-Ankle Prosthesis Towards Unified, Tuning-Free Assistance," and what they did is real time control across five different walking modes without needing those tedious manual impedance tuning or explicit intent classifiers.

Dev: That’s the headline. They managed to get an end to end system that estimates continuous actuator signals directly from onboard sensors instead of relying on a separate classifier for what the user intends.

Taro: It means they aren't tied to subject specific tuning anymore, which is a huge deal for deployment speed and scalability across different users.

Rosa: They trained this using data from eighteen people with transfemoral amputations, covering walking speeds and various ramp and stair movements. The training protocol involved a temporal convolutional network using adam optimizer with mean squared error loss.

Dev: The results show they reproduced the scaling of peak ankle torque based on walking speed during level walking, even though the able bodied participants didn't use a bypass adapter in that training set.

Taro: That scaling relationship holding up outside the training setup suggests it might generalize to users who haven't been seen before. It’s showing some real adaptability to unseen conditions.

Rosa: They also found that integrated gradients showed the model landed on biomechanically interpretable input dependencies without needing any specific architectural priors from their side.

Dev: And they tested it during stair ambulation transitions, and the key cue they learned to interpret was the user’s residual thigh lift captured by an inertial measurement unit, not vision or depth cameras.

Taro: So this is a promising step toward a unified controller that can handle context shifts across different terrains based only on mechanical sensor data.

Rosa: The final conclusion is that they achieved real time end to end control across five locomotion modes and produced mode distinct joint behaviors during ramp conditions, supporting the idea of detecting environmental context shifts.

Dev: It’s a solid piece of work showing how continuous estimation can bypass the need for those heavy manual tuning processes we’ve seen in previous systems.

Taro: But you gotta remember they have a small sample size, so we should keep an eye on how robust this generalization is when the world throws something totally unexpected at it.

Rosa: Exactly. It’s a really promising direction for making these things much more practical and deployable in the real world.

John Shim, Christoph P. O. Nuesslein, Sixu Zhou, Hanjun Kim, Kinsey Herrin, Aaron J. Young

Georgia Institute of Technology

cs.RO

Submitted: 2026-06-05

Updated: 2026-10-30

Comments: 8 pages, 7 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 93/100

The gist: Powered prostheses conventionally rely on impedance controllers that require extensive manual tuning and explicit mode classification <ref:2606.07902#pg2> In this work…

Key concepts

Impedance Controllers
These are the traditional methods used in prostheses that require extensive manual tuning by a user. They rely on explicit mode classification, meaning the system must be told exactly what action it needs to perform before it can adjust the prosthetic movement. This study aims to replace this complex setup with an automated system.
End-to-End Control
This refers to a control approach where the system takes raw sensor data and directly outputs actuator signals without intermediate, manually designed layers or classifiers. Instead of needing separate modules for sensing, classifying intent, and then controlling the motors, this method learns the entire control mapping from start to finish.
Locomotion Modes
These are distinct ways a person moves while using a prosthesis. The controller was trained to handle five specific modes: walking at different speeds (level walking), climbing ramps (ramp ascent/descent), and navigating stairs (stair ascent/descent). The system learned to switch between these behaviors based on the environment.
Temporal Convolutional Network (TCN)
This is a type of deep learning model used for training the controller. It is effective at processing time-series data, such as sensor readings over time. The TCN was trained on a dataset of locomotion data to learn the complex patterns linking sensor inputs to the required motor outputs for different walking tasks.

Terminology

Summary

Powered prostheses conventionally rely on impedance controllers that require extensive manual tuning and explicit mode classification <ref:2606. p. 007902#pg2> In this work, we present real-time deployment of an end-to-end prosthesis controller that estimates continuous actuator signals from onboard sensors, eliminating the need for intent classifiers and subjectspecific tuning <ref:2606. p. 007902#pg2> The results provide initial evidence towards end-to-end control that can provide unified, mode-adaptive prosthetic assistance without subject-specific tuning <ref:2606. p. 007902#pg2>

System Limitations Addressed

Conventional hierarchical frameworks for coordinated control face two key scalability limitations <ref:2606.07902#pg2> First, impedance controllers require extensive tuning, which can take trained clinicians up to five hours per user <ref:2606. p. 007902#pg2> Second, because impedance parameters are defined over a discrete set of locomotion modes, intent misclassification may trigger abrupt transitions in assistance <ref:2606. p. 007902#pg2> Deep learning-based control offers a promising path by generating actions directly from sensor data, which collapses the high- and mid-level hierarchy and removes the need for mode classification and FSMs <ref:2606. p. 007902#pg2>

Training Data Construction

A multi-terrain locomotion dataset was compiled from prior experiments involving the OSL under real-time impedance control <ref:2606.07902#pg2> Data from 18 individuals with TF amputation were collected, spanning level walking (LW), ramp ascent (RA) and ramp descent (RD), and stair ascent (SA) and stair descent (SD) <ref:2606. p. 1 p. 107902#pg2> The training data distribution by locomotion mode shows stride counts for stance and swing datasets <ref:2606. p. 107902#pg2>

Controller Performance on Scaling

A consistent scaling pattern was observed between realtime peak ankle plantarflexion torque and the corresponding training-data trend across walking speeds <ref:2606.07902#pg2> The deployed slope closely tracked the training scaling, indicating that the controller inferred a continuous speed context from raw sensor data and modulated assistance accordingly <ref:2606. p. 307902#pg2> Furthermore, the controller adapted to walking speed without subject-specific tuning or explicit speed estimation <ref:2606. p. 307902#pg2>

Generalization and Interpretability

The study demonstrated that the controller generalized stair transitions to unseen limb-leading conditions, producing them bilaterally when leading with either the intact-side or the prosthetic-side limb <ref:2606.07902#pg2> Integrated gradients showed the model converged on biomechanically interpretable input dependencies without architectural priors, demonstrating that end-to-end deep controllers can produce both functional and interpretable assistance <ref:2606. p. 507902#pg2> The controller produced mode-distinct joint behaviors across ramp conditions, supporting the hypothesis that it detects environmental context shifts and generates distinguishable joint behaviors across locomotion modes <ref:2606. p. 307902#pg2>

Conclusion

This study demonstrated real-time end-to-end control of a powered knee-ankle prosthesis across five locomotion modes <ref:2606.07902#pg2> Ultimately, these preliminary results represent a promising step towards a unified, tuning-free prosthesis controller that could improve deployment scalability <ref:2606. p. 507902#pg2> The results should be viewed as preliminary, pending replication in more individuals with transfemoral amputation <ref:2606. p. 507902#pg2> The small sample limits the generalization of these findings <ref:2606. p. 507902#pg2>

Pruning redundant reasoning tokens

The paper presents real-time deployment of an end-to-end prosthesis controller that estimates continuous actuator signals from onboard sensors, eliminating the need for intent classifiers and subjectspecific tuning <ref:2606.07902#pg2>

How the training signal is built

A multi-terrain locomotion dataset was compiled from prior experiments involving the OSL under real-time impedance control <ref:2606.07902#pg2> Data from 18 individuals with TF amputation were collected, spanning level walking (LW), ramp ascent (RA) and ramp descent (RD), and stair ascent (SA) and stair descent (SD) <ref:2606.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. End-to-End Control of Actuator Signals by Temporal Convolutional Networks

The system can estimate continuous actuator signals from onboard sensors, eliminating the need for intent classifiers and subject-specific tuning. This allows for real-time deployment across five locomotion modes without explicit mode classification or finite state machines.

  1. Unified Assistance Scaling Across Locomotion Modes

The controller can reproduce training data scaling relationships across environmental conditions, such as walking speed and ramp grade, even without explicit mode labels. For example, it can reproduce the training-data scaling of peak ankle torque with walking speed when excluding specific outliers traced to atypical prosthesis loading.

  1. Biomechanically Interpretable Input Dependencies

By using Integrated Gradients to attribute predictions back to sensor inputs, the AI system can identify which physical features drive its commands. This allows for a deeper understanding of the model's logic, such as confirming that My captures axial loading transmitted through the prosthesis and grows with ground reaction force, making it more robust.

  1. Continuous Gait Transition Generation

The controller can generate seamless stair transitions for both intact- and prosthetic-side-leading sequences, even when training data contained only one limb-leading sequence. This demonstrates the ability to produce complex gait transitions continuously rather than via discrete mode switching.

  1. Robustness Against Subject Variability and Input Shift

The system shows resilience against subject variability; for instance, it recovered a positive scaling slope of peak ankle torque with walking speed when excluding a specific subject's atypical loading pattern. This suggests the learned mapping generalizes to unseen users despite atypical sensor inputs like AB02’s inverted speed scaling.

Sources

Related papers