We present Kamino, a GPU-based physics solver for massively parallel simulations of heterogeneous highly-coupled mechanical systems. Implemented in Python using NVIDIA Warp and integrated into the Newton framework, it enables the application of data-driven methods, such as large-scale reinforcement learning, to complex robotic systems that exhibit strongly coupled kinematic and dynamic constraints such as kinematic loops. The latter are often circumvented by practitioners; approximating the system topology as a kinematic tree and incorporating explicit loop-closure constraints or so-called mimic joints. Kamino aims at alleviating this burden by natively supporting these types of coupling. This capability facilitates high-throughput parallelized simulations that capture the true nature of mechanical systems that exploit closed kinematic chains for mechanical advantage. Moreover, Kamino supports heterogeneous worlds, allowing for batched simulation of structurally diverse robots on a single GPU. At its core lies a state-of-the-art constrained optimization algorithm that computes constraint forces by solving the constrained rigid multi-body forward dynamics transcribed as a nonlinear complementarity problem. This leads to high-fidelity simulations that can resolve contact dynamics without resorting to approximate models that simplify and/or convexify the problem. We demonstrate RL policy training on DR Legs, a biped with six nested kinematic loops, generating a feasible walking policy while simulating 4096 parallel environments on a single GPU.

Despite recent advances in robust locomotion, bipedal robots operating in the real world remain at risk of falling. While most research focuses on preventing such events, we instead concentrate on the phenomenon of falling itself. Specifically, we aim to reduce physical damage to the robot while providing users with control over the robot's end pose. To this end, we propose a robot-agnostic reward function that balances the achievement of a desired end pose with impact minimization and the protection of critical robot parts during reinforcement learning. To make the policy robust to a broad range of initial falling conditions, and to enable the specification of an arbitrary and unseen end pose at inference time, we introduce a simulation-based sampling strategy of initial and end poses. Through simulated and real-world experiments, our work demonstrates that even bipedal robots can perform controlled, soft falls.

Behavioral Foundation Models (BFMs) offer a promising path toward universal physics-based character control by organizing a rich repertoire of physically plausible behaviors into a latent space, guided by a large-scale motion dataset. While these models excel at time-invariant tasks, such as goal-reaching and state-based reward optimization, their latent space does not directly support time-varying objectives, such as tracking a motion sequence. For tracking, existing heuristics rely on moving-window-averaging that fails to capture the nuances of highly dynamic motions. In this work, we propose a novel Latent Sequence Optimization (LSO) to address these shortcomings. Our approach combines simulation rollouts with a policy gradient update to optimize over a sequence of latents, extending the capabilities of BFMs toward precise motion tracking without requiring reward engineering and tuning. To guide the optimization toward smooth, coherent latent trajectories, we model the latent sequence using temporally correlated noise. We validate our approach across dense tracking, sparse keyframing, and direct deployment onto a real humanoid robot.

Retargeting human kinematic reference motion onto a robot's morphology remains a formidable challenge. Existing methods often produce physical inconsistencies, such as foot sliding, self-collisions, or dynamically infeasible motions, which hinder downstream imitation learning. We propose a bilevel optimization framework that jointly adapts reference motions to a robot's morphology while training a tracking policy using reinforcement learning. To make the optimization tractable, we derive an approximate gradient for the upper-level loss. Our framework requires only a sparse set of semantic rigid-body correspondences and eliminates the need for manual tuning by identifying optimal values for a parameterization expressive enough to preserve characteristic motion across different embodiments. Moreover, by integrating retargeting directly with physics simulation, we produce physically plausible motions that facilitate robust imitation learning. We validate our method in simulation and on hardware, demonstrating challenging motions for morphologies that differ significantly from a human, including retargeting onto a quadruped.

Robust state estimation for highly dynamic motion of legged robots remains challenging, especially in dynamic, contact-rich scenarios. Traditional approaches often rely on binary contact states that fail to capture the nuances of partial contact or directional slippage. This paper presents CoCo-InEKF, a differentiable invariant extended Kalman filter that utilizes continuous contact velocity covariances instead of binary contact states. These learned covariances allow the method to dynamically modulate contact confidence, accounting for more nuanced conditions ranging from firm contact to directional slippage or no contact. To predict these covariances for a set of predefined contact candidate points, we employ a lightweight neural network trained end-to-end using a state-error loss. This approach eliminates the need for heuristic ground-truth contact labels. In addition, we propose an automated contact candidate selection procedure and demonstrate that our method is insensitive to their exact placement. Experiments on a bipedal robot demonstrate a superior accuracy-efficiency tradeoff for linear velocity estimation, as well as improved filter consistency compared to baseline methods. This enables the robust execution of challenging motions, including dancing and complex ground interactions – both in simulation and in the real world.

Animated characters often move in non-physical ways and have proportions that are far from a typical walking robot. This provides an ideal platform for innovation in both mechanical design and stylized motion control. In this letter, we bring Olaf to life in the physical world, relying on reinforcement learning guided by animation references for control. To create the illusion of Olaf's feet moving along his body, we hide two asymmetric legs under a soft foam skirt. To fit actuators inside the character, we use spherical and planar linkages in the arms, mouth, and eyes. Because the walk cycle results in harsh contact sounds, we introduce additional rewards that noticeably reduce impact noise. The large head, driven by small actuators in the character's slim neck, creates a risk of overheating, amplified by the costume. To keep actuators from overheating, we feed temperature values as additional inputs to policies, introducing new rewards to keep them within bounds. We validate the efficacy of our modeling in simulation and through experiments on the robot.

Reinforcement learning (RL) has significantly advanced the control of physics-based and robotic characters that track kinematic reference motion. However, methods typically rely on a weighted sum of conflicting reward functions, requiring extensive tuning to achieve a desired behavior. Due to the computational cost of RL, this iterative process is a tedious, time-intensive task. Furthermore, for robotics applications, the weights need to be chosen such that the policy performs well in the real world, despite inevitable sim-to-real gaps. To address these challenges, we propose a multi-objective reinforcement learning framework that trains a single policy conditioned on a set of weights, spanning the Pareto front of reward trade-offs. Within this framework, weights can be selected and tuned after training, significantly speeding up iteration time. We demonstrate how this improved workflow can be used to perform highly dynamic motions with a robot character. Moreover, we explore how weight-conditioned policies can be leveraged in hierarchical settings, using a high-level policy to dynamically select weights according to the current task. We show that the multi-objective policy encodes a diverse spectrum of behaviors, facilitating efficient adaptation to novel tasks.

This technical report provides an in-depth evaluation of both established and state-of-the-art methods for simulating constrained rigid multi-body systems with hard-contact dynamics, using formulations of Nonlinear Complementarity Problems (NCPs). We are particularly interest in examining the simulation of highly coupled mechanical systems with multitudes of closed-loop bilateral kinematic joint constraints in the presence of additional unilateral constraints such as joint limits and frictional contacts with restitutive impacts. This work thus presents an up-to-date literature survey of the relevant fields, as well as an in-depth description of the approaches used for the formulation and solving of the numerical time-integration problem in a maximal coordinate setting. More specifically, our focus lies on a version of the overall problem that decomposes it into the forward dynamics problem followed by a time-integration using the states of the bodies and the constraint reactions rendered by the former. We then proceed to elaborate on the formulations used to model frictional contact dynamics and define a set of solvers that are representative of those currently employed in the majority of the established physics engines. A key aspect of this work is the definition of a benchmarking framework that we propose as a means to both qualitatively and quantitatively evaluate the performance envelopes of the set of solvers on a diverse set of challenging simulation scenarios. We thus present an extensive set of experiments that aim at highlighting the absolute and relative performance of all solvers on particular problems of interest as well as aggravatingly over the complete set defined in the suite.

Teleoperated robotic characters can perform expressive interactions with humans, relying on the operators' experience and social intuition. In this work, we propose to create autonomous interactive robots, by training a model to imitate operator data. Our model is trained on a dataset of human-robot interactions, where an expert operator is asked to vary the interactions and mood of the robot, while the operator commands as well as the pose of the human and robot are recorded. Our approach learns to predict continuous operator commands through a diffusion process and discrete commands through a classifier, all unified within a single transformer architecture. We evaluate the resulting model in simulation and with a user study on the real system. We show that our method enables simple autonomous human-robot interactions that are comparable to the expert-operator baseline, and that users can recognize the different robot moods as generated by our model. Finally, we demonstrate a zero-shot transfer of our model onto a different robotic platform with the same operator interface.

We present a constrained Rigid Body Dynamics (RBD) that guarantees satisfaction of kinematic constraints, enabling direct simulation of complex mechanical systems with arbitrary kinematic structures. To ensure constraint satisfaction, we use an implicit integration scheme. For this purpose, we derive compatible dynamic equations expressed through the quaternion time derivative, adopting an additive approach to quaternion updates instead of a multiplicative one, while enforcing quaternion unit-length as a constraint. We support all joints between rigid bodies that restrict subsets of the three translational or three rotational degrees of freedom, including position- and force-based actuation. Their constraints are formulated such that Lagrange multipliers are interpretable as joint forces and torques. We discuss a unified solution strategy for systems with redundant constraints, overactuation, and passive degrees of freedom, by eliminating redundant constraints and navigating the subspaces spanned by multipliers. As our method uses a standard additive update, we can interface with unconditionally-stable implicit integrators. Moreover, the simulation can readily be made differentiable as we show with examples.

Dynamic locomotion of legged robots is a critical yet challenging topic in expanding the operational range of mobile robots. It requires precise planning when possible footholds are sparse, robustness against uncertainties and disturbances, and generalizability across diverse terrains. Although traditional model-based controllers excel at planning on complex terrains, they struggle with real-world uncertainties. Learning-based controllers offer robustness to such uncertainties but often lack precision on terrains with sparse steppable areas. Hybrid methods achieve enhanced robustness on sparse terrains by combining both methods but are computationally demanding and constrained by the inherent limitations of model-based planners. To achieve generalized legged locomotion on diverse terrains while preserving the robustness of learning-based controllers, this paper proposes an attention-based map encoding conditioned on robot proprioception, which is trained as part of the controller using reinforcement learning. We show that the network learns to focus on steppable areas for future footholds when the robot dynamically navigates diverse and challenging terrains. We synthesized behaviors that exhibited robustness against uncertainties while enabling precise and agile traversal of sparse terrains. In addition, our method offers a way to interpret the topographical perception of a neural network. We have trained two controllers for a 12-degrees-of-freedom quadrupedal robot and a 23-degrees-of-freedom humanoid robot and tested the resulting controllers in the real world under various challenging indoor and outdoor scenarios, including ones unseen during training.

Recent advancements in generative motion models have achieved remarkable results, enabling the synthesis of lifelike human motions from textual descriptions. These kinematic approaches, while visually appealing, often produce motions that fail to adhere to physical constraints, resulting in artifacts that impede real-world deployment. To address this issue, we introduce a novel method that integrates kinematic generative models with physicsbased character control. Our approach begins by training a reward surrogate to predict the performance of the downstream non-differentiable control task, offering an efficient and differentiable loss function. This reward model is then employed to fine-tune a baseline generative model, ensuring that the generated motions are not only diverse but also physically plausible for real-world scenarios. The outcome of our processing is the Robot Motion Diffusion Model (RobotMDM), a text-conditioned kinematic diffusion model that interfaces with a reinforcement learning-based tracking controller. We demonstrate the effectiveness of this method on a challenging humanoid robot, confirming its practical utility and robustness in dynamic environments.

Recent progress in physics-based character control has made it possible to learn policies from unstructured motion data. However, it remains challenging to train a single control policy that works with diverse and unseen motions, and can be deployed to real-world physical robots. In this paper, we propose a two-stage technique that enables the control of a character with a full-body kinematic motion reference, with a focus on imitation accuracy. In a first stage, we extract a latent space encoding by training a variational autoencoder, taking short windows of motion from unstructured data as input. We then use the embedding from the time-varying latent code to train a conditional policy in a second stage, providing a mapping from kinematic input to dynamics-aware output. By keeping the two stages separate, we benefit from self-supervised methods to get better latent codes and explicit imitation rewards to avoid mode collapse. We demonstrate the efficiency and robustness of our method in simulation, with unseen user-specified motions, and on a bipedal robot, where we bring dynamic motions to the real world.

Legged robots have achieved impressive feats in dynamic locomotion in challenging unstructured terrain. However, in entertainment applications, the design and control of these robots face additional challenges in appealing to human audiences. This work aims to unify expressive, artist-directed motions and robust dynamic mobility for legged robots. To this end, we introduce a new bipedal robot, designed with a focus on character-driven mechanical features. We present a reinforcement learning-based control architecture to robustly execute artistic motions conditioned on command signals. During runtime, these command signals are generated by an animation engine which composes and blends between multiple animation sources. Finally, an intuitive operator interface enables real-time show performances with the robot. The complete system results in a believable robotic character, and paves the way for enhanced human-robot engagement in various contexts, in entertainment robotics and beyond.

The kinematic motion of a robotic character is defined by its mechanical joints and actuators that restrict the relative motion of its rigid components. Designing robots that perform a given target motion as closely as possible with a fixed number of actuated degrees of freedom is challenging, especially for robots that form kinematic loops. In this paper, we propose a technique that simultaneously solves for optimal design and control parameters for a robotic character whose design is parameterized with configurable joints. At the technical core of our technique is an efficient solution strategy that uses dynamic programming to solve for optimal state, control, and design parameters, together with a strategy to remove redundant constraints that commonly exist in general robot assemblies with kinematic loops. We demonstrate the efficacy of our approach by either editing the design of an existing robotic character, or by optimizing the design of a new character to perform a desired motion.

Dynamic locomotion in rough terrain requires accurate foot placement, collision avoidance, and planning of the underactuated dynamics of the system. Reliably optimizing for such motions and interactions in the presence of imperfect and often incomplete perceptive information is challenging. We present a complete perception, planning, and control pipeline, which can optimize motions for all degrees of freedom of the robot in real time. To mitigate the numerical challenges posed by the terrain, a sequence of convex inequality constraints is extracted as local approximations of foothold feasibility and embedded into an online model-predictive controller. Steppability classification, plane segmentation, and a signed distance field are precomputed per elevation map to minimize the computational effort during the optimization. A combination of multiple-shooting, real-time iteration, and a filter-based line search is used to solve the formulated problem reliably and at high rate. We validate the proposed method in scenarios with gaps, slopes, and stepping stones in simulation and experimentally on the ANYmal quadruped platform, resulting in state-of-the-art dynamic climbing.

Legged robots are designed to perform highly dynamic motions. However, it remains challenging for users to retarget expressive motions onto these complex systems. In this paper, we present a Differentiable Optimal Control (DOC) framework that facilitates the transfer of rich motions from either animals or animations onto these robots. Interfacing with either motion capture or animation data, we formulate retargeting objectives whose parameters make them agnostic to differences in proportions and numbers of degrees of freedom between input and robot. Optimizing these parameters over the manifold spanned by optimal state and control trajectories, we minimize the retargeting error. We demonstrate the utility and efficacy of our modeling by applying DOC to a Model-Predictive Control (MPC) formulation, showing retargeting results for a family of robots of varying proportions and mass distribution. With a hardware deployment, we further show that the retargeted motions are physically feasible, while MPC ensures that the robots retain their capability to react to unexpected disturbances.

We describe an optimization-based framework to perform complex locomotion skills for robots with legs and wheels. The generation of complex motions over a long-time horizon often requires offline computation due to current computing constraints and is mostly accomplished through trajectory optimization (TO). In contrast, model predictive control (MPC) focuses on the online computation of trajectories, robust even in the presence of uncertainty, albeit mostly over shorter time horizons and is prone to generating nonoptimal solutions over the horizon of the task’s goals. Our article’s contributions overcome this trade-off by combining offline motion libraries and online MPC, uniting a complex, long-time horizon plan with reactive, short-time horizon solutions. We start from offline trajectories that can be, for example, generated by TO or sampling-based methods. Also, multiple offline trajectories can be composed out of a motion library into a single maneuver. We then use these offline trajectories as the cost for the online MPC, allowing us to smoothly blend between multiple composed motions even in the presence of discontinuous transitions. The MPC optimizes from the measured state, resulting in feedback control, which robustifies the task’s execution by reacting to disturbances and looking ahead at the offline trajectory. With our contribution, motion designers can choose their favorite method to iterate over behavior designs offline without tuning robot experiments, enabling them to author new behaviors rapidly. Our experiments demonstrate complex and dynamic motions on our traditional quadrupedal robot ANYmal and its roller-walking version. Moreover, the article’s findings contribute to evaluating five planning algorithms.

Modern robotic systems are endowed with superior mobility and mechanical skills that make them suited to be employed in real-world scenarios, where interactions with heavy objects and precise manipulation capabilities are required. For instance, legged robots with high payload capacity can be used in disaster scenarios to remove dangerous material or carry injured people. It is thus essential to develop planning algorithms that can enable complex robots to perform motion and manipulation tasks accurately. In addition, online adaptation mechanisms with respect to new, unknown environments are needed. In this work, we impose that the optimal state-input trajectories generated by Model Predictive Control (MPC) satisfy the Lyapunov function criterion derived in adaptive control for robotic systems. As a result, we combine the stability guarantees provided by Control Lyapunov Functions (CLFs) and the optimality offered by MPC in a unified adaptive framework, yielding an improved performance during the robot's interaction with unknown objects. We validate the proposed approach in simulation and hardware tests on a quadrupedal robot carrying un-modeled payloads and pulling heavy boxes.

The ability to generate dynamic walking in real-time for bipedal robots with input constraints and underactuation has the potential to enable locomotion in dynamic, complex and unstructured environments. Yet, the high-dimensional nature of bipedal robots has limited the use of full-order rigid body dynamics to gaits which are synthesized offline and then tracked online. In this work we develop an online nonlinear model predictive control approach that leverages the full-order dynamics to realize diverse walking behaviors. Additionally, this approach can be coupled with gaits synthesized offline via a desired reference to enable a shorter prediction horizon and rapid online re-planning, bridging the gap between online reactive control and offline gait planning. We demonstrate the proposed method, both with and without an offline gait, on the planar robot AMBER-3M in simulation and on hardware.