Rapid co-design of Buoyancy-assisted robots for Challenging Locomotion using Gaussian Evolutionary Specialists
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: "Rapid co-design of Buoyancy-assisted robots for Challenging Locomotion using Gaussian Evolutionary Specialists".
Dev: Here's the gist of the paper. Designing high-performance legged robots requires jointly optimizing morphology and control <ref:2606.07424#pg2> The resulting specialists are integrated into a design sampling loop, replacing costly re-training with direct evaluation <ref:2606.07424#pg6>.
Rosa: This framework discovers designs with 5 − 25% higher performance than naive universal policies on the Buoyancy-Assisted Light Legged Unit (BALLU) <ref:2606.07424#pg2> On hardware, a GES optimized design overcomes a 24 cm tall obstacle - 3× improvement over the baseline BALLU design <ref:2606.07424#pg2>.
Dev: The Core Problem Addressed Robot co-design jointly optimizes morphology and control to achieve task performance that neither component can reach in isolation <ref:2606.07424#pg2> The primary bottleneck in this bi-level optimization problem is the cost of repeated controller training per candidate design <ref:2606.07424#pg2>.
Rosa: Prior work using universal policies fails because they suffer from behavioral diversity collapse, converging to a single strategy that performs sub-optimally across designs <ref:2606.07424#pg2>.
Paper summary: Rosa: So we're looking at this paper about "Rapid co-design of Buoyancy-assisted robots for Challenging Locomotion using Gaussian Evolutionary Specialists," and it’s about how you can find the best physical shape for a robot that uses balloons to float, while also finding the best way to control it.
Dev: Right, so the core idea here is tackling this really hard bi-level optimization problem where you have to optimize both the robot's physical structure and its control policy at the same time, which neither component can do alone <ref:2606.07424#pg2>.
Taro: And the main thing they’re fighting is that if you use standard universal policies, they just collapse into one single strategy that doesn't work well for different robot shapes <ref:2606.07424#pg2>.
Rosa: Exactly, and end-to-end Mixture-of-Experts models also struggle because their representation collapses too <ref:2606.07424#pg2>.
Dev: So this framework, Gaussian Evolutionary Specialists or GES, aims to fix that by separating how the design space is partitioned from how the policies are learned <ref:2606.07424#pg2>.
Taro: They use this two-stage approach where specialists train on specific regions of the design space based on some Gaussian clusters, and then they refine those designs iteratively <ref:2606.07424#pg2>.
Rosa: The goal of this whole setup is to replace the costly process of training a new controller for every single robot design with just direct evaluation by these specialized policies <ref:2606.07424#pg6>.
Dev: When they tested it on the Buoyancy-Assisted Light Legged Unit or BALLU, they found designs that are five to twenty-five percent better than those naive universal policies <ref:2606.07424#pg2>.
Taro: And on hardware, this optimized design can clear a twenty four centimeter tall obstacle, which is three times better than the original baseline design <ref:2606.07424#pg2>.
Rosa: It also cuts down the time it takes to co-design by thirty seven percent <ref:2606.07424#pg2>.
Dev: That reduction in time is pretty significant, but we have to remember that they are dealing with a very nonlinear system on the BALLU platform threenine <ref:2606.07424#pg1>.
Taro: And the dynamics of those balloon-based robots introduce these severe aerodynamic nonlinearities, like buoyancy and drag forces that are hard to model analytically <ref:2606.07424#pg3>.
Rosa: So, how does this GES framework actually work to handle those tough dynamics when you're trying to find the perfect leg length or stiffness setting?
Dev: Stage one of the GES process involves iterative Gaussian territory evolution where specialists train on designs sampled from their specific Gaussian cluster <ref:2606.07424#pg2>.
Taro: And the border designs, those in between clusters, are competitively assigned to the best-performing specialist until all the territories collectively tile the entire design space <ref:2606.07424#pg2>.
Rosa: That sounds like they are systematically exploring all parts of that design space efficiently without having to test every possibility individually <ref:2606.07424#pg1>.
Dev: Stage two is where the specialists actually become the evaluators, and they serve as a zero-shot evaluator for design optimization, which replaces that repeated per-candidate reinforcement learning training <ref:2606.07424#pg2>.
Taro: This means you feed a new design into all these trained specialists, and they tell you which one performs best on that specific shape <ref:2606.07424#pg3>.
Rosa: So, the big picture here is that this GES framework resolves the problem of behavioral diversity collapse by decoupling the partitioning of the design space from the policy learning itself <ref:2606.07424#pg1>.
Dev: It’s interesting because it specifically addresses both failures, so you get better performance on obstacle traversal and ramp climbing in both two-dimensional and three-dimensional design spaces <ref:2606.07424#pg3>.
Taro: The analysis also showed that higher Gaussian cluster radii and higher spectral performance coefficients consistently lead to better results when walking up a ramp <ref:2606.07424#pg2>.
Rosa: And structurally, the way GES assigns designs to its dominant specialist creates a hard partition over the design space, which is different from how end-to-end Mixture-of-Experts handles it <ref:2606.07424#pg3>.
Dev: From an engineering standpoint, they reduced the total co-design wall-clock time from about one hundred hours down to sixty three hours, which is a thirty seven percent reduction over using per-design specialist Bayesian optimization <ref:2606.07424#pg2>.
Taro: That speedup is significant because it makes brutal force design exploration much more practical when dealing with these complex systems <ref:2606.07424#pg1>.
Rosa: So, to wrap up the findings from "Rapid co-design of Buoyancy-assisted robots for Challenging Locomotion using Gaussian Evolutionary Specialists," this research shows that by using this GES framework, you can jointly optimize the physical shape and the control policy much more effectively <ref:2606.07424#pg1>.
Dev: It means for people designing these kinds of robots, you don't have to train a whole new controller from scratch every time you change a leg length or stiffness setting anymore <ref:2606.07424#pg2>.
Taro: The implication is that this approach could be extended beyond the BALLU platform to other types of robots like quadrupeds or humanoids because the framework seems platform agnostic <ref:2606.07424#pg1>.
Rosa: It’s a lot of work, but it suggests a path forward where we can move toward more robust, dynamically optimized legged systems without getting bogged down in those massive retraining costs <ref:2606.07424#pg1>.
Conclusion: Rosa: So we’re wrapping up this look at "Rapid co-design of Buoyancy-assisted robots for Challenging Locomotion using Gaussian Evolutionary Specialists."
Dev: Yeah, so basically, it’s about how you can find the perfect physical shape for a buoyant robot and its control system at the same time without spending forever re-training everything.
Taro: Right, they used this Gaussian Evolutionary Specialists framework to tackle that bi-level optimization problem.
Rosa: They found that this method lets you discover designs with five to twenty-five percent better performance than just using naive universal policies for these buoyancy robots.
Dev: And on the physical hardware, they showed that one of these optimized designs can clear a twenty four centimeter obstacle, which is three times better than the baseline design they started with.
Taro: It’s interesting because it addresses two major problems at once: the behavioral diversity collapse and the representation collapse you see in end-to-end mixture of experts models.
Rosa: That’s right, they propose this GES framework which separates design space partitioning from policy learning to capture all those diverse behaviors explicitly.
Dev: The main contribution is how they use iterative Gaussian territory evolution to replace costly per-candidate controller training with direct evaluation during the design loop.
Taro: It makes the whole co-design process much faster, cutting down on wall-clock time by thirty seven percent compared to other methods they tested.
Rosa: So it’s about making robot co-design practical for complex platforms like buoyant robots where you can't just test every shape by training a brand new controller each time.
Dev: It means that for future work, they are looking at an adaptive version of this system that can grow or merge those Gaussian territories based on how well they cover the design space and how good the performance is.
Taro: And they’re also thinking about making it platform agnostic so it could be used for other things like quadruped robots or even humanoids.
Rosa: That’s what we'll be looking at next—how this approach to co-design might translate when you move away from just a simple light wheeled unit and into much more complex robot bodies.
Ankit Sinha, Nitish Sontakke, Dennis Hong, Yusuke Tanaka, Sehoon Ha
Georgia Institute of Technology · University of California, Los Angeles · ETH Zurich
cs.RO
Submitted: 2026-06-05
Updated: 2026-10-30
Comments: Under review
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 94/100
The gist: Designing high-performance legged robots requires jointly optimizing morphology and control <ref:2606.07424#pg2> The resulting specialists are integrated into a design sampling loop…
Key concepts
- Robot Co-design
- This is the process of simultaneously optimizing two interdependent parts of a robot: its physical shape (morphology) and the control software that dictates how it moves. The goal is to find a combination where neither part can reach its best performance in isolation, leading to better overall task success.
- Behavioral Diversity Collapse
- This occurs when universal policies, which try to work for many designs at once, converge on a single movement strategy. This collapse means the policy fails to capture the diverse optimal behaviors available across different robot shapes, resulting in poor performance for most designs.
- Gaussian Evolutionary Specialists (GES)
- GES is a two-stage framework that explicitly captures diverse behaviors by assigning each design region to a specialized expert policy. Stage 1 involves specialists training on specific Gaussian regions of the design space, and Stage 2 uses these trained specialists as fast evaluators for new designs.
- Mixture-of-Experts (MoE)
- MoE is an architecture where multiple specialized neural networks (experts) work together. In this context, it's used to handle the complexity of robot control by assigning different parts of the design space to different experts, which helps prevent all experts from converging to the same suboptimal behavior.
Terminology
Summary
Designing high-performance legged robots requires jointly optimizing morphology and control <ref:2606. p. 0 Rapid co-design of Buoyancy-assisted robots for Challenging
Locomotion using Gaussian Evolutionary Specialists
Ankit Sinha1
, Nitish Sontakke1
, Dennis Hong2
, Yusuke Tanaka3
and Sehoon Ha1
Abstract— Designing high-performance legged robots requires jointly optimizing morphology and control. Model-free
Reinforcement Learning (RL) offers an alternative to modelpredictive control for developing robust controllers without
explicitly specifying robot dynamics. GES assigns specialist policies to evolving Gaussian
regions and iteratively refines them via training, probing, and
territory expansion. The resulting specialists are integrated into
a design sampling loop, replacing costly re-training with direct
evaluation. GES assigns specialist policies to evolving Gaussian
regions and iteratively refines them via training, probing, and
territory expansion. The resulting specialists are integrated into
a design sampling loop, replacing costly re-training with direct
evaluation. The resulting specialists are integrated into
a design sampling loop, replacing costly re-training with direct
evaluation. When tested on the Buoyancy-Assisted Light Legged
Unit (BALLU), GES discovers designs with 5 − 25% higher
performance than naive universal policies. When tested on the Buoyancy-Assisted Light Legged
Unit (BALLU), GES discovers designs with 5 − 25% higher
performance than naive universal policies. On hardware, a
GES optimized design overcomes a 24 cm tall obstacle - 3×
improvement over the baseline BALLU design. When tested on the Buoyancy-Assisted Light Legged
Unit (BALLU), GES discovers designs with 5 − 25% higher
performance than naive universal policies. On hardware, a
GES optimized design overcomes a 24 cm tall obstacle - 3×
improvement over the baseline BALLU design.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The Core Problem Addressed
Robot co-design jointly optimizes morphology and control to achieve task performance that neither component can reach in isolation <ref:2606.07424#pg2> The primary bottleneck in this bi-level optimization problem is the cost of repeated controller training per candidate design <ref:2606. p. 0 approaches such as differentiable co-optimization [6] and
meta-learning [7] has reduced this cost, but repeated controller optimization per candidate remains the bottleneck. The cumulative cost of repeated controller training makes
brutal-force design exploration impractical. While RL has shown success in
locomotion, using it in the co-design inner loop is expensive
due to repeated policy training. Universal policies conditioned
on morphology offer a promising alternative, but suffer from
behavioral diversity collapse, converging to a single strategy that
performs sub-optimally across designs. Mixture-of-Experts
(MoE) [16] is a natural remedy, assigning each region of the
design space to a dedicated specialist to capture multi-modal
behaviors. However, end-to-end MoE training can fail due
to the representation collapse problem [17]: the MoE router
arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
The GES Framework Mechanism
Gaussian Evolutionary Specialists (GES) is a two-stage framework that decouples design-space partitioning from policy learning to capture diverse behaviors explicitly <ref:2606.07424#pg2>. Stage 1 involves iterative Gaussian territory evolution where specialists train on designs sampled from their associated Gaussian cluster, and the territories collectively tile the design space <ref:2606. p. 1 In Stage 1, multiple specialists are assigned a Gaussian
territory: each specialist trains on designs sampled from
its territory, border designs are competitively assigned to
the best-performing specialist, and each Gaussian is refit to
its won designs - repeating until the territories collectively
tile. In Stage 2, the trained specialists serve as a zero-shot
evaluator for design optimization, substituting per-candidate
RL training. F candidate designs are sampled near the 2Σ borders
of all Gaussian territories. Each candidate is evaluated by all
K specialists, and assigned to the specialist that achieves the
highest task performance score J on that design:
k
∗ = arg max
k
J(πk, c), ηk∗ ← ηk∗ ∪ {c}.Where the paper says it
Where the paper says it
Performance Gains and Analysis
GES outperforms both baselines across tasks and design space dimensionalities <ref:2606.07424#pg2> For obstacle traversal in the 3D design space, GES finds designs clearing 54 cm against 41 cm for e2e MLP, a peak gain of 31. p. 4 On head-tohead comparison, GES wins on 94% of designs against e2e
MLP and 72% against e2e MoE. In the 3D design space,
GES again finds designs clearing 54 cm against 41 cm for
e2e MLP, a peak gain of 31. Mixture-of-Experts
(MoE) [16] is a natural remedy, assigning each region of the
design space to a dedicated specialist to capture multi-modal
behaviors. However, end-to-end MoE training can fail due
to the representation collapse problem [17]: the MoE router
arXiv:2606. Design Analysis. Higher GCR and higher SPCF consistently
produce better performance: high buoyancy reduces effective
weight to enable flight, and high spring stiffness stores the
potential energy needed for jumping. In GES, each design is assigned to the specialist
whose Gaussian territory covers that region, producing a hard
partition in which every specialist dominates its subspace. This structural difference explains the ∼20% degradation of
e2e MoE as design-space increases from 2D to 3D.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Computational Efficiency and Hardware Validation
GES reduces total co-design time from ∼100 h to ∼63 h, a 37% reduction over per-design specialist BO on the 3D design space <ref:2606.07424#pg2>. The framework is inherently parallelizable, reducing Stage 1 complexity from O(KN) to O(N) when K GPUs are used <ref:2606. p. 6 GES reduces total co-design time
from ∼100 h to ∼63 h, a 37% reduction over per-design
specialist BO on the 3D design space (RTX 2070 Super,
single GPU). GES is inherently parallelizable: training and
probing each specialist are independent, so K GPUs reduce
Stage 1 complexity from O(KN) to O(N). 5% mean improvement over a monolithic universal policy on obstacle traversal. On hardware,
the co-designed BALLU clears a 24 cm obstacle, a 3×
improvement over the baseline design, and GES reduces codesign wall-clock time by 37%.Where the paper says it
Where the paper says it
Conclusion and Future Directions
GES resolves behavioral diversity collapse by decoupling design-space partitioning from expert specialization via iterative Gaussian territory evolution <ref:2606.07424#pg2>. Three directions remain open for future work, including developing an adaptive variant that grows or merges territories based on coverage and performance <ref:2606. p. 7 Three directions remain open for future work. First, GES
currently assumes a fixed number of specialists K; an
adaptive variant that grows or merges territories based on
coverage and performance would reduce this hyperparameter
sensitivity. First, GES
currently assumes a fixed number of specialists K; an
adaptive variant that grows or merges territories based on
coverage and performance would reduce this hyperparameter
sensitivity. Second, while we validate on a single buoyancyassisted platform, the GES framework is platform-agnostic
and should generalize to other robots like quadrupeds and
humanoids. Third, GES assumes continuous and bounded
design space. In future, we plan to re-design this framework
by relaxing the above assumptions.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Rapid co-design of Buoyancy-assisted robots for Challenging Locomotion using Gaussian Evolutionary Specialists Ankit Sinha, Nitish Sontakke, Dennis Hong, Yusuke Tanaka and Sehoon Ha Abstract— Designing high-performance legged robots requires jointly optimizing morphology and control <ref:2606.07424#pg2> Model-free Reinforcement Learning (RL) offers an alternative to modelpredictive control for developing robust controllers without explicitly specifying robot dynamics <ref:2606. p. 0 Rapid co-design of Buoyancy-assisted robots for Challenging
Locomotion using Gaussian Evolutionary Specialists
Ankit Sinha1
, Nitish Sontakke1
, Dennis Hong2
, Yusuke Tanaka3
and Sehoon Ha1
Abstract— Designing high-performance legged robots requires jointly optimizing morphology and control. Model-free
Reinforcement Learning (RL) offers an alternative to modelpredictive control for developing robust controllers without
explicitly specifying robot dynamics. Model-free
Reinforcement Learning (RL) offers an alternative to modelpredictive control for developing robust controllers without
explicitly specifying robot dynamics. Thus, we have seen the
use of RL to train controllers and evaluate designs for robot
morphology optimization. Thus, we have seen the
use of RL to train controllers and evaluate designs for robot
morphology optimization. While RL has shown success in
locomotion, using it in the co-design inner loop is expensive
due to repeated policy training. While RL has shown success in
locomotion, using it in the co-design inner loop is expensive
due to repeated policy training. Universal policies conditioned
on morphology offer a promising alternative, but suffer from
behavioral diversity collapse, converging to a single strategy that
performs sub-optimally across designs. Universal policies conditioned
on morphology offer a promising alternative, but suffer from
behavioral diversity collapse, converging to a single strategy that
performs sub-optimally across designs. On the other hand, endto-end Mixture-of-Experts (MoE) architectures fail due to a
collapse in its representation. On the other hand, endto-end Mixture-of-Experts (MoE) architectures fail due to a
collapse in its representation. We propose Gaussian Evolutionary Specialists (GES), a framework that decouples design-space
partitioning from policy learning to capture diverse behaviors
explicitly. We propose Gaussian Evolutionary Specialists (GES), a framework that decouples design-space
partitioning from policy learning to capture diverse behaviors
explicitly. GES assigns specialist policies to evolving Gaussian
regions and iteratively refines them via training, probing, and
territory expansion. The resulting specialists are integrated into
a design sampling loop, replacing costly re-training with direct
evaluation. When tested on the Buoyancy-Assisted Light Legged
Unit (BALLU), GES discovers designs with 5 − 25% higher
performance than naive universal policies. On hardware, a
GES optimized design overcomes a 24 cm tall obstacle - 3×
improvement over the baseline BALLU design. Moreover, GES
curtails design optimization time by 37%. For instance, optimizing link lengths
in a robotic arm expands the reachable workspace for a
target task, while tuning leg proportions in a legged robot
shifts the trade-off between energy efficiency and agility [2]. The Buoyancy-Assisted Light Legged Unit (BALLU) [3],
[4] studied in this work is one such platform: its aerodynamic sensitivity to balloon inflation level and leg proportions means that design choices have an out-sized effect
on task performance. The Buoyancy-Assisted Light Legged Unit (BALLU) [3],
[4] studied in this work is one such platform: its aerodynamic sensitivity to balloon inflation level and leg proportions means that design choices have an out-sized effect
on task performance. Moreover, BALLU’s thin links, nonlinear spring-servo actuation, and aerodynamic modeling
inaccuracies in simulation make multi-parameter design optimization challenging. However, co-design is an intrinsically
bi-level optimization problem: the outer loop must evaluate
hundreds of morphology candidates, and each evaluation
requires training a controller from scratch [2], [5]. Prior
1Ankit, Nitish and Sehoon are with the School of Interactive Computing,
Georgia Institute of Technology, TSRB 85 5th St NW, Atlanta, GA - 30332,
USA {asinha389,nitishsontakke,sha9}@gatech. approaches such as differentiable co-optimization [6] and
meta-learning [7] has reduced this cost, but repeated controller optimization per candidate remains the bottleneck. The cumulative cost of repeated controller training makes
brutal-force design exploration impractical. This cost grows substantially when reinforcement learning
(RL) is used for the inner loop [1], [8], as each policy optimization relies on millions of environment interactions [2],
[9]. Prior work has addressed this through graph-based
design search [10], [11], but these methods still require percandidate RL training and preserve the core computational
bottleneck. One promising approach is to pretrain a universal
policy conditioned on morphology parameters, so that each
outer-loop trial reduces to a zero-shot evaluation [2], [12],
[13]. This formulation decouples controller learning from
design search, and makes large-scale co-design feasible. This formulation decouples controller learning from
design search, and makes large-scale co-design feasible. In practice, however, universal policies fail as co-design
evaluators. In practice, however, universal policies fail as co-design
evaluators. A design-conditioned universal policy converges
to a single locomotion strategy due to conflicting gradients from diverse morphologies [14], [15], as elaborated in
Sec. A design-conditioned universal policy converges
to a single locomotion strategy due to conflicting gradients from diverse morphologies [14], [15], as elaborated in
Sec. A collapsed evaluator cannot distinguish performance differences across designs, causing the outer optimization loop to lose its search signal. Mixture-of-Experts
(MoE) [16] is a natural remedy, assigning each region of the
design space to a dedicated specialist to capture multi-modal
behaviors. However, end-to-end MoE training can fail due
to the representation collapse problem [17]: the MoE router
arXiv:2606. Mixture-of-Experts
(MoE) [16] is a natural remedy, assigning each region of the
design space to a dedicated specialist to capture multi-modal
behaviors. However, end-to-end MoE training can fail due
to the representation collapse problem [17]: the MoE router
arXiv:2606. Once homogenized,
the routing signals are insufficient to drive specialization. We propose Gaussian Evolutionary Specialists (GES), a
two-stage framework that aims to address both failure modes. We propose Gaussian Evolutionary Specialists (GES), a
two-stage framework that aims to address both failure modes. In Stage 1, multiple specialists are assigned a Gaussian
territory: each specialist trains on designs sampled from
its territory, border designs are competitively assigned to
the best-performing specialist, and each Gaussian is refit to
its won designs - repeating until the territories collectively
tile. In Stage 1, multiple specialists are assigned a Gaussian
territory: each specialist trains on designs sampled from
its territory, border designs are competitively assigned to
the best-performing specialist, and each Gaussian is refit to
its won designs - repeating until the territories collectively
tile. In Stage 2, the trained specialists serve as a zero-shot
evaluator for design optimization, substituting per-candidate
RL training. In Stage 2, the trained specialists serve as a zero-shot
evaluator for design optimization, substituting per-candidate
RL training. We validate GES on BALLU [3], [9], which dynamics
is highly nonlinear, making co-design both necessary and
challenging. We validate GES on BALLU [3], [9], which dynamics
is highly nonlinear, making co-design both necessary and
challenging. GES outperforms both baselines across obstacle
traversal and ramp climbing tasks in 2D and 3D design
spaces, with up to 25. GES outperforms both baselines across obstacle
traversal and ramp climbing tasks in 2D and 3D design
spaces, with up to 25. 5% mean improvement over a monolithic universal policy on obstacle traversal. 5% mean improvement over a monolithic universal policy on obstacle traversal. On hardware,
the co-designed BALLU clears a 24 cm obstacle, a 3×
improvement over the baseline design, and GES reduces codesign wall-clock time by 37%. 5% mean improvement over a monolithic universal policy on obstacle traversal. On hardware,
the co-designed BALLU clears a 24 cm obstacle, a 3×
improvement over the baseline design, and GES reduces codesign wall-clock time by 37%. On hardware,
the co-designed BALLU clears a 24 cm obstacle, a 3×
improvement over the baseline design, and GES reduces codesign wall-clock time by 37%. Our main contributions are:
• We identify behavioral diversity collapse as a previously unrecognized failure mode of universal policies
in robot co-design. Our main contributions are:
• We identify behavioral diversity collapse as a previously unrecognized failure mode of universal policies
in robot co-design. • We propose GES, which resolves this via iterative Gaussian territory evolution and addresses the representation
collapse problem with end-to-end MoEs.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
A. Robot Co-design Robot co-design has been approached through several paradigms <ref:2606. p. 1 RELATED WORK
A. Robot Co-design
Robot co-design has been approached through several
paradigms. Robot Co-design
Robot co-design has been approached through several
paradigms. Differentiable frameworks [6], [18], [19], [20],
[21], [22] jointly optimize morphology and control via analytical gradients, but require differentiable simulators. Differentiable frameworks [6], [18], [19], [20],
[21], [22] jointly optimize morphology and control via analytical gradients, but require differentiable simulators. Metalearning [7] enables rapid adaptation to new designs, but still
requires per-candidate fine-tuning at query time. Metalearning [7] enables rapid adaptation to new designs, but still
requires per-candidate fine-tuning at query time. Graph-based
methods [10], [11], [23] efficiently explore combinatorial
design spaces, but scale poorly to continuous parameters. Graph-based
methods [10], [11], [23] efficiently explore combinatorial
design spaces, but scale poorly to continuous parameters. Evolutionary methods [1], [24] offer good diversity, but
require many evaluations to converge. Evolutionary methods [1], [24] offer good diversity, but
require many evaluations to converge. Recent generative
approaches [25] use diffusion models to synthesize robot
morphologies, broadening the search beyond predefined parameterizations. Recent generative
approaches [25] use diffusion models to synthesize robot
morphologies, broadening the search beyond predefined parameterizations. All of these share a common bottleneck:
each candidate design requires an independent controller
optimization. All of these share a common bottleneck:
each candidate design requires an independent controller
optimization. Strgar and Kriegman [26] amortize this cost
via universal policy pretraining, but restrict their study to
soft robots and report population-level diversity collapse. Strgar and Kriegman [26] amortize this cost
via universal policy pretraining, but restrict their study to
soft robots and report population-level diversity collapse. Bjelonic et al. Bjelonic et al. [2] are the closest to our work: they train a
design-conditioned policy for quadruped joint optimization,
but do not examine its failure modes as a design evaluator. [2] are the closest to our work: they train a
design-conditioned policy for quadruped joint optimization,
but do not examine its failure modes as a design evaluator. Our work identifies behavioral diversity collapse as one such
failure mode and proposes GES to address it. Universal Policies for Locomotion
Learning locomotion policies that generalize across different robot morphologies has emerged as a research direction
KNEE
FEMUR
TIBIA
PELVIS
FOOT
BALLOON
HIP
Fig. 2: BALLU Hardware
toward universal robot control. 2: BALLU Hardware
toward universal robot control. Prior work has progressed
from modular GNN-based approaches [27], to transformerbased controllers [28], to simple MLP policies trained via
morphology randomization that deploys zero-shot across
diverse platforms [12]. Prior work has progressed
from modular GNN-based approaches [27], to transformerbased controllers [28], to simple MLP policies trained via
morphology randomization that deploys zero-shot across
diverse platforms [12]. [29], [30] show that universal
policies can adapt to unknown dynamics through online
system identification and meta-learning. More recent multiembodiment frameworks achieve large-scale generalization
through attention and diffusion-based training [31], [32]. We extend this universal control paradigm to accelerate
the co-design of rigid body robots. We extend this universal control paradigm to accelerate
the co-design of rigid body robots. We find that a monolithic
policy [2], [12] trained with morphology randomization converges to uni-modal solutions and fails to capture the diverse
optimal behaviors across the design space. We find that a monolithic
policy [2], [12] trained with morphology randomization converges to uni-modal solutions and fails to capture the diverse
optimal behaviors across the design space. As a result, such a
policy is an unreliable evaluator for co-design. As a result, such a
policy is an unreliable evaluator for co-design. This motivates
our mixture-of-specialists approach that preserves behavioral
diversity across the full design space. This motivates
our mixture-of-specialists approach that preserves behavioral
diversity across the full design space. Buoyancy-assisted Robots
Buoyancy-assisted robots use helium balloons [33] or
pneumatic inflation [34] to counteract gravity, offering intrinsic safety for human-robot interaction. Buoyancy-assisted Robots
Buoyancy-assisted robots use helium balloons [33] or
pneumatic inflation [34] to counteract gravity, offering intrinsic safety for human-robot interaction. BALLU [3], [4] uses
helium inflation to enable bipedal locomotion with cabledriven actuation and minimal hardware mass. BALLU [3], [4] uses
helium inflation to enable bipedal locomotion with cabledriven actuation and minimal hardware mass. However, the
aerodynamics of balloon-based robots pose challenges for
control and sim-to-real transfer. However, the
aerodynamics of balloon-based robots pose challenges for
control and sim-to-real transfer. Sontakke et al. Sontakke et al. [9] address
this through residual physics learning and system identification, and demonstrate successful deployment of RL policies
on hardware. [9] address
this through residual physics learning and system identification, and demonstrate successful deployment of RL policies
on hardware. Our work extends beyond basic locomotion
to co-optimizing BALLU for obstacle traversal and ramp
walking, which demands more dynamic motion.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 3 ---
A. BALLU Platform BALLU [3], [4] is a bipedal robot that uses helium-filled balloons to offset the majority of its weight <ref:2606. p. 1 PROBLEM FORMULATION
A. BALLU Platform
BALLU [3], [4] is a bipedal robot that uses helium-filled
balloons to offset the majority of its weight. BALLU Platform
BALLU [3], [4] is a bipedal robot that uses helium-filled
balloons to offset the majority of its weight. The platform
(Fig. The platform
(Fig. 2) consists of five rigid links: a pelvis and two upper
and lower limbs connected by spring-loaded knee joints Its lightweight
construction and fall-safe operation make it well-suited for
human-facing applications like environmental monitoring
and social robotics [3], [4]. Despite its mechanical simplicity, BALLU presents significant control challenges. Despite its mechanical simplicity, BALLU presents significant control challenges. The helium balloons introduce
severe aerodynamic nonlinearities: buoyancy and drag forces
that are difficult to model analytically and vary with balloon
inflation level and environmental conditions [3]. The helium balloons introduce
severe aerodynamic nonlinearities: buoyancy and drag forces
that are difficult to model analytically and vary with balloon
inflation level and environmental conditions [3]. These dynamics make model-free RL the most practical approach to
controller design, as demonstrated by Sontakke et al. Mixture-of-Experts
(MoE) [16] is a natural remedy, assigning each region of the
design space to a dedicated specialist to capture multi-modal
behaviors. However, end-to-end MoE training can fail due
to the representation collapse problem [17]: the MoE router
arXiv:2606. These dynamics make model-free RL the most practical approach to
controller design, as demonstrated by Sontakke et al. Design Space
The BALLU morphology design space, D, is parameterized by four variables: femur length lf , tibia length lt,
gravity compensation ratio GCR (ratio of buoyancy force to
robot weight), and spring stiffness, SPCF. 45 m
Tibia length lt 0. ratio GCR 0. At GCR = 1. 0
the robot is effectively weightless; lower values increase the
effective gravitational load on the legs. 0
the robot is effectively weightless; lower values increase the
effective gravitational load on the legs. The SPCF governs
knee joint compliance, which directly affects energy storage
during dynamic maneuvers. The parameter bounds and the
baseline BALLU designs [3] are summarized in Table I. In our experiments, we evaluate GES on two design
parameter space: D2 = {(GCR, SPCF) : lf = ¯lf , lt =
¯lt} ⊂ R
2
and D3 = {(GCR, SPCF, l) : lf = lt = l} ⊂ R
3
. Bi-level Co-design Objective
The goal of robot co-design is to jointly find a morphology
θm and a control policy θc that maximize task performance:
max
θm, θc
J(θc, θm). (1)
This joint optimization is intractable to solve directly due to
the interdependency between morphology and control. (1)
This joint optimization is intractable to solve directly due to
the interdependency between morphology and control. We
therefore decompose it into a bi-level form:
max
θm
J(θ
∗
c
, θm), θ
∗
c = arg max
θc
J(θc, θm), (2)
where the inner loop finds the optimal controller for a given
morphology and the outer loop searches over the morphology
space to maximize the resulting performance. MDP Formulation
We model locomotion control as a Markov Decision
Process (MDP) defined by the tuple (S, A, T , R, γ). The
proprioceptive state st ∈ S comprises the robot’s base linear
and angular velocity, joint positions, joint velocities, and the
last action. The augmented state s˜t = [st, θm] appends the
design vector. The action at ∈ R
2
specifies joint position
targets for the left and right knees, which are tracked by
a PD controller that converts targets to joint torques. The action at ∈ R
2
specifies joint position
targets for the left and right knees, which are tracked by
a PD controller that converts targets to joint torques. The
policy πθc
(at | s˜t) is parameterized as a Gaussian MLP and
optimized via PPO to maximize expected discounted return:
max
θc
E
"X
T
t=0
γ
t
r(st, at)
#
(3)
Reward functions. The
policy πθc
(at | s˜t) is parameterized as a Gaussian MLP and
optimized via PPO to maximize expected discounted return:
max
θc
E
"X
T
t=0
γ
t
r(st, at)
#
(3)
Reward functions. For obstacle traversal, the reward combines a navigation term, a forward velocity term, and a
jumping term that encourages toe clearance:
robs = w1 (−∥g − s∥
2
2
)
| {z }
rnav
+w2 (vx)
|{z}
rvel
+w3 (e
c·ztoe − 1)
| {z }
rjump
, (4)
where g is the goal position, s is the robot’s current base position, vx is the forward velocity, and ztoe = min(z
L
toe, zR
toe)
is the minimum toe height across both legs.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
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
-
Improve robot co-design by replacing costly per-candidate RL training with direct evaluation using Gaussian Evolutionary Specialists (GES). This allows for design optimization without repeated policy re-training, curtailing design optimization time by 37 percent.
-
Implement GES to overcome behavioral diversity collapse in universal policies by decoupling design-space partitioning from policy learning through iterative Gaussian territory evolution and end-to-end Mixture-of-Experts (MoE) architectures. This enables the system to capture multi-modal behaviors explicitly rather than converging to a single, suboptimal strategy.
-
Enable zero shot evaluation of designs by using GES density router for Bayesian Optimization (BO) based co-design. This router selects the nearest specialist based on Gaussian likelihood, replacing expensive inner loop RL with a fast performance score J hat that updates the BO surrogate belief over the design space D.
Abstract
Designing high-performance legged robots requires jointly optimizing morphology and control. Model-free reinforcement learning (RL) enables learning complex controllers during co-design, but repeatedly training policies across candidate morphologies makes design optimization computationally expensive. Existing approaches can amortize policy learning across morphologies, but a single shared policy tends to converge toward one control strategy regardless of morphology, which we refer to as behavioral diversity collapse. We introduce Gaussian Evolutionary Specialists (GES), an RL-based co-design framework that maintains a population of specialist policies whose design-space territories evolve according to their relative control performance. We evaluate GES on the Buoyancy-Assisted Light Legged Unit (BALLU) robot for obstacle traversal and ramp walking over design spaces of up to four dimensions. GES achieves higher best design performance than Monolithic and Mixture-of-Experts baselines, with gains of 5.0-9.0 cm in obstacle height and 10.0-12.7 points in ramp grade. For the same 200 design-optimization evaluations, GES reduces wall-clock co-design time by 67.4% compared with retraining a policy for every candidate. Finally, teleoperated hardware experiments evaluate the optimized BALLU morphology, which traverses obstacles of up to 24 cm, compared with 8 cm on the original design.
Sources
- Multi-Loco: Unifying Multi-Embodiment Legged Locomotion via Reinforcement Learning Augmented Diffusion
- Recon: Reducing Conflicting Gradients from the Root for Multi-Task Learning
- Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning
- Proximal Policy Optimization Algorithms
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation