Learning and planning in structured latent spaces for legged robot locomotion
<p>Quadruped robots are capable of operating in human-oriented environments littered with stairs, no-step areas and cabling, whilst carrying multiple sensor payloads. Therefore, these robots are popular in the industrial, maintenance and energy sectors. However, the versatility of these robots...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: |
| _version_ | 1811141292047990784 |
|---|---|
| author | Mitchell, AL |
| author2 | Posner, H |
| author_facet | Posner, H Mitchell, AL |
| author_sort | Mitchell, AL |
| collection | OXFORD |
| description | <p>Quadruped robots are capable of operating in human-oriented environments littered with stairs, no-step areas and cabling, whilst carrying multiple sensor payloads. Therefore, these robots are popular in the industrial, maintenance and energy sectors. However, the versatility of these robots comes with the cost of complex nonlinear dynamics. These dynamics have both continuous and discrete elements, for example, joint-space motion such as joint angles and torques vary smoothly, but contact states are discontinuous and non-differentiable. Optimising trajectories constrained by these dynamics can be complex and slow resulting in narrow basins of convergence and limited windows of operation. Indeed, these methods often require precomputed footstep locations and schedules to be real-time capable meaning that these gait characteristics cannot be varied on demand. Model-free methods tend to be black-box and it is not possible to vary the resulting policies or inspect the properties of the resulting gaits without some sort of prior inductive bias. An example of such an inductive bias is the relative phase of each leg normally referred to as the gait phase. Locomotion is periodic since each leg continuously makes and breaks contact and the gait phase describes the contact and swing relationships between the feet.</p>
<p>An ideal planner would have the following properties: capable of a diverse range of motions or gaits, and integrate proprioceptive and perceptive information such as vision for terrain traversal all while being robust and interpretable. Indeed in nature, we see a diverse range of robust skills such as continuous gait switching and modulation to external stimuli for example speed demands or changes in terrains. This thesis investigates whether such a controller can be created by learning a unified representation for quadruped locomotion across skills. We train a generative model namely a variational autoencoder (VAE) on examples of locomotion trajectories. In doing so, we discover that the robot’s gait characteristics (footswing length and height) become disentangled into separate dimensions of the latent space. As a result, these parameters can be varied independently by altering the amplitudes of two oscillatory signals in latent space. Furthermore, training the model with examples of distinct gait types (trot, crawl and pace) reveals that the latent space is able to infer meaningful correlations across the gaits. As a result, the VAE is able to infer novel and unseen gait transitions between the motions. With the addition of the rudimentary terrain encoding, the latent space is able to adjust its structure to adapt to changes in terrain. A learnt planner in the latent space adjusts the gait characteristics in response to the terrain and improves the robustness of foothold location during the climbing phase. At this point, we show that a latent space forms a suitable planner capable of delivering a diverse range of locomotion skills all while being robust and interpretable. Since our methods are inspired by observations of nature, we posit if there exists similar representations in the natural world. In prior work by Churchland et al, it was discovered that both locomotion and reaching skills are encoded by the circular firing of neurons in the motor cortex of primates. These circular neural dynamics are strikingly similar to those in the locomotion latent space so far as the time periods and phase lag between the cycles in the motor cortex and latent space are of the same magnitude. Inspired by the similarities between the locomotion and manipulation oscillations in the motor cortex, we train a manipulation-specific latent space and discover that the dimensions of this space are also meaningfully disentangled. Longitudinal and lateral movements of the robot’s end effector map to separate latent-space dimensions. Therefore, rotational dynamics similar to those in the motor cortex can be used to solve reaching tasks deployed on the real robot manipulator.</p>
<p>This thesis explores the opportunities that a unified latent representation of locomotion can provide and shows how it is used to create a versatile planner. In doing so, a diverse range of motions with continuously adjustable gait characteristics is achievable. Furthermore, a latent space trained with examples of multiple distinct gaits is able to infer the gait phase across gaits. Exploration of this space reveals continuous gait transitions via novel contact schedules. The addition of a rudimentary terrain encoding and a learnt planner in latent space permits the robust traversal of uneven terrains. The planner is deployed on the real robot and all experiments are analysed using data from runs using the real platform.</p> |
| first_indexed | 2024-09-25T04:35:33Z |
| format | Thesis |
| id | oxford-uuid:bc6cb8f8-8f4a-4ae1-a538-7f7a90b828d6 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:35:33Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | oxford-uuid:bc6cb8f8-8f4a-4ae1-a538-7f7a90b828d62024-09-20T08:20:47ZLearning and planning in structured latent spaces for legged robot locomotionThesishttp://purl.org/coar/resource_type/c_db06uuid:bc6cb8f8-8f4a-4ae1-a538-7f7a90b828d6RoboticsControlEnglishHyrax Deposit2024Mitchell, ALPosner, HHavoutis, IUmenberger, Jvan der Smagt, P<p>Quadruped robots are capable of operating in human-oriented environments littered with stairs, no-step areas and cabling, whilst carrying multiple sensor payloads. Therefore, these robots are popular in the industrial, maintenance and energy sectors. However, the versatility of these robots comes with the cost of complex nonlinear dynamics. These dynamics have both continuous and discrete elements, for example, joint-space motion such as joint angles and torques vary smoothly, but contact states are discontinuous and non-differentiable. Optimising trajectories constrained by these dynamics can be complex and slow resulting in narrow basins of convergence and limited windows of operation. Indeed, these methods often require precomputed footstep locations and schedules to be real-time capable meaning that these gait characteristics cannot be varied on demand. Model-free methods tend to be black-box and it is not possible to vary the resulting policies or inspect the properties of the resulting gaits without some sort of prior inductive bias. An example of such an inductive bias is the relative phase of each leg normally referred to as the gait phase. Locomotion is periodic since each leg continuously makes and breaks contact and the gait phase describes the contact and swing relationships between the feet.</p> <p>An ideal planner would have the following properties: capable of a diverse range of motions or gaits, and integrate proprioceptive and perceptive information such as vision for terrain traversal all while being robust and interpretable. Indeed in nature, we see a diverse range of robust skills such as continuous gait switching and modulation to external stimuli for example speed demands or changes in terrains. This thesis investigates whether such a controller can be created by learning a unified representation for quadruped locomotion across skills. We train a generative model namely a variational autoencoder (VAE) on examples of locomotion trajectories. In doing so, we discover that the robot’s gait characteristics (footswing length and height) become disentangled into separate dimensions of the latent space. As a result, these parameters can be varied independently by altering the amplitudes of two oscillatory signals in latent space. Furthermore, training the model with examples of distinct gait types (trot, crawl and pace) reveals that the latent space is able to infer meaningful correlations across the gaits. As a result, the VAE is able to infer novel and unseen gait transitions between the motions. With the addition of the rudimentary terrain encoding, the latent space is able to adjust its structure to adapt to changes in terrain. A learnt planner in the latent space adjusts the gait characteristics in response to the terrain and improves the robustness of foothold location during the climbing phase. At this point, we show that a latent space forms a suitable planner capable of delivering a diverse range of locomotion skills all while being robust and interpretable. Since our methods are inspired by observations of nature, we posit if there exists similar representations in the natural world. In prior work by Churchland et al, it was discovered that both locomotion and reaching skills are encoded by the circular firing of neurons in the motor cortex of primates. These circular neural dynamics are strikingly similar to those in the locomotion latent space so far as the time periods and phase lag between the cycles in the motor cortex and latent space are of the same magnitude. Inspired by the similarities between the locomotion and manipulation oscillations in the motor cortex, we train a manipulation-specific latent space and discover that the dimensions of this space are also meaningfully disentangled. Longitudinal and lateral movements of the robot’s end effector map to separate latent-space dimensions. Therefore, rotational dynamics similar to those in the motor cortex can be used to solve reaching tasks deployed on the real robot manipulator.</p> <p>This thesis explores the opportunities that a unified latent representation of locomotion can provide and shows how it is used to create a versatile planner. In doing so, a diverse range of motions with continuously adjustable gait characteristics is achievable. Furthermore, a latent space trained with examples of multiple distinct gaits is able to infer the gait phase across gaits. Exploration of this space reveals continuous gait transitions via novel contact schedules. The addition of a rudimentary terrain encoding and a learnt planner in latent space permits the robust traversal of uneven terrains. The planner is deployed on the real robot and all experiments are analysed using data from runs using the real platform.</p> |
| spellingShingle | Robotics Control Mitchell, AL Learning and planning in structured latent spaces for legged robot locomotion |
| title | Learning and planning in structured latent spaces for legged robot locomotion |
| title_full | Learning and planning in structured latent spaces for legged robot locomotion |
| title_fullStr | Learning and planning in structured latent spaces for legged robot locomotion |
| title_full_unstemmed | Learning and planning in structured latent spaces for legged robot locomotion |
| title_short | Learning and planning in structured latent spaces for legged robot locomotion |
| title_sort | learning and planning in structured latent spaces for legged robot locomotion |
| topic | Robotics Control |
| work_keys_str_mv | AT mitchellal learningandplanninginstructuredlatentspacesforleggedrobotlocomotion |