Koopman Model-Based Reinforcement Learning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Koopman Model-Based Reinforcement Learning".
Title and authors: Rosa: So, we’re looking at this paper called "Koopman Model-Based Reinforcement Learning". It proposes an algorithm for controlling nonlinear robotic systems when you don't know how they move, by combining a deep Koopman model with a critic function to figure out the best control policy.
Dev: That’s right. The title itself points to the core idea: using Koopman theory, which lets us describe complex dynamics in a linear way, combined with reinforcement learning techniques for control. It’s about tackling those unknown nonlinear dynamics on physical robots that we can't just model perfectly from scratch.
Taro: I always wonder how this handles the uncertainty when the robot actually starts moving outside of simulation, Rosa. If the underlying dynamics are nonlinear and unknown, how reliable is this approach in a real-world setting?
Rosa: Well, what they do is learn a stochastic Koopman operator based learning method to approximate those nonlinear dynamics using a deep neural network lifting function. This lifts the unknown system into this linear representation, which then allows them to compute the policy gradient much more efficiently than traditional methods.
Dev: That efficiency comes from the fact that this linear structure simplifies the model-dependent policy gradient computation down to just a constant matrix product, which is pretty fast for online updates on physical hardware. But they also address a major issue with long simulations by using one-step model propagation instead of multi-step rollouts.
Taro: One step propagation sounds safer for real-time control because it limits how far the prediction error can accumulate in a single loop, but does that mean we lose some of the necessary information to plan ahead over a long horizon?
Rosa: That’s exactly what they try to avoid. They combine that one-step model propagation with a critic function that estimates the remaining future cost based on what the model predicts for just the next state, which keeps track of long-term cost information without letting prediction errors pile up too much.
Dev: And they frame their policy objective by combining the immediate cost you’re paying right now with this state-action critic estimate of that future cost at the one-step predicted state. It’s a clever way to balance short-term action with long-term planning without needing massive model rollouts every time.
Title and authors: Taro: So, they are using this critic to compensate for the fact that the Koopman model is only a one-step predictor, rather than trying to solve the entire trajectory at once? That sounds like a necessary trade-off when dealing with unknown dynamics.
Rosa: Exactly. It’s presented as the first method to integrate stochastic deep Koopman learning with state-action critic based policy optimization, which is a significant step in how we approach model-based reinforcement learning for complex physical systems.
Dev: Theoretically, they establish convergence properties under standard stochastic approximation assumptions. Lemma one shows that the parameters of the deep Koopman model converge to their optimal values as k goes to infinity, and lemma two shows that the critic parameter converges to its optimal value associated with the policy you’re evolving.
Taro: As long as those convergence lemmas hold in practice, it gives us a strong theoretical foundation for trusting that these learned models and critics will eventually lead to a stable control policy. What about the practical deployment?
Rosa: They show strong sample efficiency on tasks like HalfCheetah and quadruped locomotion when compared against classical model-free and model-based reinforcement learning baselines, which suggests they learn much faster in practice. Even on physical hardware, like a quadruped robot performing a challenging sequence of maneuvers, the results are promising.
Dev: They also note that on physical robots, like the Unitree Go1 quadruped, this approach achieves higher final returns and faster convergence than SAC or RKL while keeping the computational cost relatively low per policy update. That real-time performance is what keeps me interested in it for control loops.
Taro: The hardware validation is important because simulation can be very misleading when you introduce real-world factors like actuator asymmetry or unmodeled dynamics that we have to deal with constantly outside of the lab. How does this method handle those specific physical challenges?
Rosa: They tested it on a physical quadruped robot performing bipedal walking, and the robot successfully performed straight walking as well as left and right turning maneuvers while maintaining an upright posture during that test. It shows a level of resilience to these kinds of real-world difficulties.
Dev: The computational efficiency is also a big deal for deployment because the linear structure of the deep Koopman model lets them compute policy gradients much faster, requiring less computation time per update than RKL or MPC-optimal solutions which solve the MBRL problem directly.
Title and authors: Taro: So, to summarize for someone just listening to this—this paper is about taking nonlinear systems we don't understand and finding a way to control them efficiently by using a linear approximation from the Koopman operator, paired with a critic that manages the long-term cost estimation while keeping things computationally light enough for real robots.
Rosa: That’s right. It tackles two big headaches: learning complex nonlinear dynamics and preventing prediction errors from exploding during long rollouts. The convergence guarantees they provide under their stated assumptions add a layer of mathematical rigor to the practical results we see on the hardware trials.
Dev: I think what this means for us is that we might be able to deploy more capable model-based controllers on robots that are currently too complex or too slow because of the computational expense associated with traditional model-based methods. It’s about making model-based learning practical for real-time control loops.
Taro: I’m interested in where they stop working, though. The paper notes that deployment introduces challenges like actuator asymmetry and hardware protection mechanisms, which means reliable real-world control still heavily depends on how well the data covers those specific deployment factors.
Rosa: So, for future work, they plan to look into things like safetyaware learning and better replay buffer sampling to make it even more robust when we move this out of controlled environments. It’s clear they see the path forward involves making these models safer and more adaptable.
Dev: Indeed, the computational savings become increasingly significant as the algorithm iteration grows, which is crucial if we want to use this for continuous, online policy improvement where latency matters a lot. I think this Koopman Model-Based Reinforcement Learning approach offers a solid path forward for real-world nonlinear control.
Taro: It does show promise for complex locomotion tasks when you compare it against other methods like SAC or EMBRL on the HalfCheetah and quadruped tests, suggesting it’s competitive in terms of performance while being faster to converge.
Rosa: So, we’ve looked at how this paper tackles nonlinear dynamics with a Koopman model, how they use a critic to manage prediction errors without needing long rollouts, and how the resulting method shows strong sample efficiency and hardware applicability on quadruped robots like the Unitree Go1. That’s our rundown of "Koopman Model-Based Reinforcement Learning".
The paper's summary: Rosa: So, we're looking at the summary of this paper on Koopman Model-Based Reinforcement Learning for controlling nonlinear robots with unknown dynamics and how that actually works in practice.
Dev: Basically, it’s about using a deep Koopman model to turn that messy nonlinear robot movement into something linear so the computer can handle it efficiently.
Rosa: Right. They're combining this Koopman model with a critic function to figure out the best control policy by looking ahead at the future cost. It tries to get the long-term plan without getting lost in prediction errors over long time steps.
Dev: That’s exactly the trick they use to keep things stable during those long simulations. They take what you can predict in just one step and combine it with a critic estimate of everything that comes after that point.
Rosa: And the results show this approach is much more sample efficient than older methods like model-free or even some other model-based ones on tasks like walking robots. It learns the control policy faster in practice.
Dev: On the hardware side, they tested it on a real quadruped robot and found that it actually performs complex maneuvers, like turning and maintaining balance while walking upright, which is pretty impressive for real-world application.
Rosa: They also show that the computational cost per update is much lower than other methods because of how the linear structure simplifies the policy gradient calculation. It’s faster for real-time loops.
Dev: The theoretical part of it is solid too; they prove that as you run the learning process, both the Koopman model parameters and that critic function settle down to their best possible values asymptotically.
Rosa: So what this means for us is that we can finally start deploying more sophisticated model-based controllers on robots today instead of just in simulation, because it’s faster and more sample efficient.
Dev: It points toward a way to make these complex learning algorithms practical for real-time control loops where you need low latency. But the authors also flag that real hardware introduces issues like uneven actuators or unmodeled dynamics that still need careful data coverage to handle.
Rosa: They’re looking ahead to safetyaware learning and better ways of sampling data from the replay buffer because deploying this outside a clean lab environment still comes with real challenges.
Dev: So, while the theory is strong and the hardware validation is promising, it's clear that making this work reliably in every messy real-world scenario is still an open question.
The paper's improvements: Tom: So, we’re talking about the specific improvements the authors suggest for this Koopman Model-Based Reinforcement Learning paper and what those changes mean for future work.
Rosa: The main thing they suggest is to look at safetyaware learning, which means building in extra checks to make sure the robot doesn't do anything dangerous when things go wrong.
Dev: That’s a good point because the current setup relies on convergence under certain assumptions; adding safety constraints might change those assumptions.
Rosa: They also talk about improving how data is sampled from the replay buffer, which is important because if you sample poorly, the AI won't learn what it needs to learn effectively.
Dev: That relates to robustness in deployment; if we can get better data quality from the experience we gather, the resulting policy should be more reliable when it hits a real robot.
Rosa: And they’re also thinking about scaling this up, looking at how this works for larger robotic systems instead of just small quadruped robots.
Dev: Scaling is a big hurdle because as you get bigger, the computational demands on that linear Koopman representation can explode if it doesn't scale correctly.
Rosa: So they’re essentially trying to find ways to make the learning process more robust and scalable so we can move beyond these small test setups.
Dev: It’s about bridging that gap between simulation performance and reliable real-world deployment, which is where the real engineering challenge lies.
Conclusion: Rosa: So we're wrapping up on this paper, "Koopman Model-Based Reinforcement Learning," by summarizing how it tackles unknown nonlinear robot dynamics and long prediction errors in model-based reinforcement learning.
Dev: It basically shows that you can use a deep Koopman model to make the dynamics linear, which lets you compute the policy gradient much faster than traditional methods.
Rosa: And they used a critic function to estimate the future cost at one step, which helps keep those long-term plans stable and avoids prediction errors piling up.
Dev: The convergence guarantees they established for both the model and that critic are pretty solid under standard assumptions, so we have a decent theoretical foundation here for trusting that it will eventually work.
Taro: I still have to ask what happens when the robot encounters something completely outside of its training data during a real-world run; how gracefully does this system recover?
Rosa: They admit that deployment introduces challenges like uneven actuators and unmodeled physics, so they’re planning future work around making it safer and more adaptable.
Dev: That makes sense because the current setup relies on those assumptions holding true during the learning process, but real life is rarely perfectly clean.
Taro: So it’s less about solving every physical problem perfectly right now and more about building a framework that can handle the mess better than existing methods.
Rosa: Exactly; they’re looking at safetyaware learning and better data sampling to push this model out into the real world reliably.
Dev: It’s a solid piece of work showing how to make model-based control more practical for physical systems without needing massive computation time per update.
Rosa: We've seen strong sample efficiency and good performance on locomotion tasks, so it looks like a viable path forward for controlling complex robots in the future.
Dev: This Koopman Model-Based Reinforcement Learning paper gives us a clear way to handle those nonlinear dynamics on hardware that we can actually deploy today with reasonable computation overhead.
Wenjian Hao, Paulo C Heredia, Shaoshuai Mou
Purdue University
cs.RO, cs.SY, eess.SY
Submitted: 2026-11-10
Updated: 2026-11-10
Code: https://github.com/unitreerobotics/unitree_sdk2
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 92/100
The gist: Koopman model-based reinforcement learning (K-MBRL) is proposed as an algorithm for controlling nonlinear robotic systems with unknown dynamics by combining a deep Koopman model with a critic…
Key concepts
- Koopman Model
- A deep neural network that approximates complex nonlinear system dynamics by transforming them into a linear space. This allows the robot's movement to be modeled using simpler, more efficient linear algebra operations instead of complex nonlinear calculations.
- Critic Function
- A component that estimates the remaining future cost of a state-action pair. It works alongside the model to guide policy learning by providing an accurate estimate of how good a particular action is in the long run, helping to correct errors from short-term predictions.
- Policy Objective Approximation
- The goal function used to train the robot's control policy. K-MBRL approximates this objective by mixing the immediate cost with a future cost estimate derived from the critic and a one-step prediction from the Koopman model, balancing short-term action with long-term goals.
Terminology
Summary
Koopman model-based reinforcement learning (K-MBRL) is proposed as an algorithm for controlling nonlinear robotic systems with unknown dynamics by combining a deep Koopman model with a critic function to approximate the policy objective, which offers computational efficiency and mitigates prediction errors. p. 0 Heredia 1
, Shaoshuai Mou1
Abstract
We propose a Koopman model-based reinforcement learning (K-MBRL) algorithm for controlling nonlinear robotic
systems with unknown dynamics. The K-MBRL learns a control policy by minimizing an approximate policy objective
that combines the analytical immediate cost with an estimate of the future cost evaluated at the one-step predicted state.Where the paper says it
Core Methodology
The K-MBRL algorithm learns a control policy by minimizing an approximate policy objective that combines the analytical immediate cost with an estimate of the future cost evaluated at the one-step predicted state. This is achieved through two main components: developing a stochastic Koopman operator-based learning method to approximate unknown nonlinear dynamics in a linear representation, and introducing a critic function that estimates the remaining future cost. p. 0 To this end, we develop a stochastic Koopman operator-based learning method to approximate the unknown nonlinear
dynamics in a linear representation, while a critic function estimates the remaining future cost. This formulation offers
two key advantages. This formulation offers
two key advantages. First, the linear Koopman dynamics simplify the model-dependent policy gradient computation to
a constant matrix product, enabling computationally efficient policy improvement.Where the paper says it
Where the paper says it
Policy Objective Approximation
The algorithm approximates the policy objective by combining the analytical immediate cost with a state-action critic estimate of the future cost evaluated at the one-step state predicted by the learned deep Koopman model (DKM). The formulation aims to mitigate the accumulation of model prediction errors associated with long-horizon rollouts while retaining long-term cost information through the critic. p. 0 First, the linear Koopman dynamics simplify the model-dependent policy gradient computation to
a constant matrix product, enabling computationally efficient policy improvement. Second, one-step model propagation
mitigates the accumulation of prediction errors associated with long-horizon rollouts while retaining long-term cost
information through the critic. The algorithm approximates the policy objective by
combining the analytical immediate cost with a stateaction critic estimate of the future cost evaluated at the
one-step state predicted by the learned deep Koopman
model (DKM). To the best of our knowledge, this is
the first method to integrate stochastic DKM learning
with state-action critic-based policy optimization.Where the paper says it
Where the paper says it
Theoretical Guarantees
Policy convergence guarantees are established under standard stochastic approximation assumptions. Lemma 1 establishes the convergence of the DKM parameters in (9) to their optimal values, and Lemma 2 establishes the convergence of the critic parameter qk in (11) to the optimal critic parameter associated with the evolving policy wk. p. 5 Lemma 1 establishes the convergence of the DKM
parameters in (9) to their optimal values, the proof of which
is done by extending standard SGD analysis in (Robbins and
Monro 1951). Lemma 2 establishes the convergence of the
critic parameter qk in (11) to the optimal critic parameter
associated with the evolving policy wk, which is based on
two-timescale convergence analysis for actor-critic methods
in (Konda and Borkar 1999). Theorem 1 establishes that the proposed K-MBRL algorithm approaches first-order stationarity of the approximate
policy objective J. Although this does not guarantee convergence of the policy πw∗ to π
∗
in (1), simulations in Fig.Where the paper says it
Where the paper says it
Experimental Validation
Simulations on HalfCheetah and quadruped four-legged locomotion tasks demonstrate strong sample efficiency and improved performance compared with classical model-free and model-based reinforcement learning baselines. On a physical quadruped robot performing a challenging sequence of maneuvers, the proposed K-MBRL is further validated. p. 0 A pendulum swing-up task is used to assess policy gradient approximation accuracy and conduct ablation
studies on the learning mechanism and key hyperparameters. The proposed K-MBRL is further validated on a physical
quadruped robot performing a challenging sequence of maneuvers: rising to an upright posture supported by two
rounded rear feet, maintaining balance during bipedal walking, and safely returning to the ground, as shown in Fig. In summary, Figs. 3, 7, and 8 demonstrate that K-MBRL
achieves higher final returns and faster convergence than
SAC, RKL, and EMBRL across the considered tasks, while
maintaining relatively low computational cost.Where the paper says it
Where the paper says it
Hardware Applicability
The proposed K-MBRL is evaluated on a physical Unitree Go1 quadruped robot performing bipedal walking to improve the quadruped robot’s resilience to front-leg failures or enable locomotion with its front legs available for other tasks. The experiments show that the robot successfully performs straight walking as well as left and right turning maneuvers while maintaining the upright bipedal posture. p. 0 A pendulum swing-up task is used to assess policy gradient approximation accuracy and conduct ablation
studies on the learning mechanism and key hyperparameters. The proposed K-MBRL is further validated on a physical
quadruped robot performing a challenging sequence of maneuvers: rising to an upright posture supported by two
rounded rear feet, maintaining balance during bipedal walking, and safely returning to the ground, as shown in Fig. In summary, Figs. 3, 7, and 8 demonstrate that K-MBRL
achieves higher final returns and faster convergence than
SAC, RKL, and EMBRL across the considered tasks, while
maintaining relatively low computational cost.Where the paper says it
Where the paper says it
Computational Efficiency
The linear structure of the DKM facilitates policy gradient computation, reducing computation time per policy update. The results show that K-MBRL requires much less computation time at each policy update than RKL and MPC-optimal since both of them directly employ MPC to solve the MBRL problem. p. 6 3 shows that K-MBRL, SAC, and RKL successfully
solve the MBRL problem as well as MPC-optimal, whereas
EMBRL exhibits higher episode costs, potentially due
to the accumulation of dynamics prediction errors over
long prediction horizons. Although the proposed K-MBRL
converges a little bit slower than RKL, K-MBRL requires
much less computation time at each policy update than RKL
and MPC-optimal since both of them directly employ MPC
to solve the MBRL. By further
increasing the number of neurons per layer in the DNNbased dynamics models to be double of those in K-MBRL,
the large DNN-based variant K-MBRL: DNN (larger)
achieves average testing return closer to the K-MBRL, at
the expense of much longer computation times for policy
updates because of the increased model complexity. These
results highlight the advantage of using the deep Koopman
model in achieving competitive control performance with
fewer model parameters and lower computation time per
policy update, with the computational savings becoming
increasingly significant as the number of iterations grows.Where the paper says it
Where the paper says it
Model Flexibility
The K-MBRL framework is flexible regarding system dynamics learning, as it can be adapted to approximate the unknown dynamics using a deep Koopman model (DKM), a DNN-based dynamics model, or known structures. The paper shows that the proposed K-MBRL achieves similar convergence speed and control performance to its adaptation to given dynamics structure fµ. p. 5 4 of Section 5. 1, the proposed
K-MBRL achieves similar convergence speed and control
performance to its adaptation to given dynamics structure fµ. 4, the proposed K-MBRL and its variants
always achieves control policies in this pendulum example
that are close to the MPC-optimal one, no matter whether
the dynamics learning part employs a DKM, DNN, or known
structures. However, the K-MBRL does not require prior
knowledge of the system structure.Where the paper says it
Where the paper says it
Hardware Challenges
Hardware deployment introduces additional challenges associated with actuator asymmetry, hardware protection mechanisms, and unmodeled dynamics. These considerations highlight the importance of data coverage and deployment-specific factors for reliable real-world robotic control. p. 3 The red blocks highlight the approximate policy
objective mechanism. Red and blue arrows show the policy
learning flow, and black arrows show the policy deployment flow. Bipedal walking on the Go1 is particularly challenging
because the robot is not mechanically designed for this mode
of locomotion. Unlike humanoid robots, whose feet typically
provide relatively large and approximately planar contact
surfaces, the Go1 has small, approximately spherical feet. A pendulum swing-up task is used to assess policy gradient approximation accuracy and conduct ablation
studies on the learning mechanism and key hyperparameters. The proposed K-MBRL is further validated on a physical
quadruped robot performing a challenging sequence of maneuvers: rising to an upright posture supported by two
rounded rear feet, maintaining balance during bipedal walking, and safely returning to the ground, as shown in Fig. Bipedal walking on the Go1 is particularly challenging
because the robot is not mechanically designed for this mode
of locomotion. Unlike humanoid robots, whose feet typically
provide relatively large and approximately planar contact
surfaces, the Go1 has small, approximately spherical feet.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The proposed K-MBRL addresses two key challenges in model-based policy optimization: learning complex nonlinear robotic dynamics and limiting the accumulation of prediction errors during long-horizon model propagation. The resulting computational savings become increasingly significant as algorithm iteration grows. p. 9 By further
increasing the number of neurons per layer in the DNNbased dynamics models to be double of those in K-MBRL,
the large DNN-based variant K-MBRL: DNN (larger)
achieves average testing return closer to the K-MBRL, at
the expense of much longer computation times for policy
updates because of the increased model complexity. These
results highlight the advantage of using the deep Koopman
model in achieving competitive control performance with
fewer model parameters and lower computation time per
policy update, with the computational savings becoming
increasingly significant as the number of iterations grows. • Policy convergence guarantees of K-MBRL are
provided in the presence of coupled dynamics and
critic updates under a fixed-dataset setting. The
analysis further characterizes how dynamics and
critic approximation errors affect policy convergence,
providing theoretical insights into policy learning
under both perfect and imperfect approximations. Performance of the K-MBRL
with constant step sizes is very close to the K-MBRL with
Robbins-Monro step sizes, as shown in Fig. 5 Simulations
In this section, we evaluate the proposed K-MBRL in Algorithm 1 in terms of control performance, sample efficiency,
and computational efficiency by detailed comparisons with
existing RL methods.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
-
Policy Objective Approximation: The system can minimize a combined objective by
combining the analytical immediate cost with a state-action critic estimate of the future cost evaluated at the one-step predicted state,
which mitigates prediction errors while retaining long-term information through a critic function. -
Dynamics Learning: The system can approximate unknown nonlinear dynamics using a deep Koopman model (DKM), which learns
a linear representation in a lifted space
via a deep neural network lifting function, allowing forcomputationally efficient policy improvement
through constant matrix product computation. -
Policy Optimization Efficiency: Policy gradient computation is simplified because the linear DKM facilitates the process, enabling
real-time control through feedforward evaluation without online optimization.
-
Error Mitigation: The algorithm mitigates prediction error accumulation over long rollouts by using
one-step model propagation
combined with a critic function that estimates the remaining future cost. -
Convergence Guarantees: Theoretical analysis provides convergence properties, establishing that under certain assumptions, the DKM parameters and critic parameters
converge to their optimal values asymptotically as k goes to infinity.
-
Sample Efficiency: The proposed algorithm demonstrates
strong sample efficiency
on tasks like HalfCheetah and quadruped locomotion compared with classical model-free and model-based reinforcement learning baselines. -
Hardware Applicability: The system can be deployed on physical robots, such as the Unitree Go1 quadruped, enabling complex maneuvers like
rising to an upright posture supported by two rounded rear feet
during bipedal walking.
Abstract
We propose a Koopman model-based reinforcement learning (K-MBRL) algorithm for controlling nonlinear robotic systems with unknown dynamics. The K-MBRL learns a control policy by minimizing an approximate policy objective that combines the analytical immediate cost with an estimate of the future cost evaluated at the one-step predicted state. To this end, we develop a stochastic Koopman operator-based learning method to approximate the unknown nonlinear dynamics in a linear representation, while a critic function estimates the remaining future cost. This formulation offers two key advantages. First, the linear Koopman dynamics simplify the model-dependent policy gradient computation to a constant matrix product, enabling computationally efficient policy improvement. Second, one-step model propagation mitigates the accumulation of prediction errors associated with long-horizon rollouts while retaining long-term cost information through the critic. The learned policy enables real-time control through feedforward evaluation without online optimization. We establish policy learning convergence guarantees under standard stochastic approximation assumptions. Simulations on HalfCheetah and quadruped four-legged locomotion tasks demonstrate strong sample efficiency and improved performance compared with classical model-free and model-based reinforcement learning baselines. A pendulum swing-up task is used to assess policy gradient approximation accuracy and conduct ablation studies on the learning mechanism and key hyperparameters. The proposed K-MBRL is further validated on a physical quadruped robot performing a challenging sequence of maneuvers: rising to an upright posture supported by two rounded rear feet, maintaining balance during bipedal walking, and safely returning to the ground, as shown in Fig.1.
Sources
- Model-Augmented Actor-Critic: Backpropagating through Paths
- Playing Atari with Deep Reinforcement Learning
- Koopman global linearization of contact dynamics for robot locomotion and manipulation enables elaborate control
- Sample-Efficient Online Control Policy Learning with Real-Time Recursive Model Updates
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