Latest / Embodied AI 101 / Get Up, Adapt, Survive: Recovery Skills for Small Quadrupeds

Transcript

Get Up, Adapt, Survive: Recovery Skills for Small Quadrupeds · Embodied AI 101

• Get Up, Adapt, Survive: Recovery Skills for Small Quadrupeds. Real-to-Sim-to-Real: Learning Agile and Robust Recovery Skills With Terrain Imagination via Adversarial Imitation • Tangyu Qian, Huayang Yin, Zhen Kan • University of Science and Technology of China • IEEE Robotics and Automation Letters • 2026. Research note. As of August 1, 2026, the authors’ public project repository still labels the downloadable paper as “coming soon,” despite the article’s publication in IEEE Robotics and Automation Letters. This discussion therefore distinguishes documented claims—from the paper abstract, project materials, experiment videos, and deployment information—from engineering interpretation. In particular, I will not invent numerical table entries or hyperparameters that are not publicly exposed. Recovery Is Not Just Locomotion in Reverse. A field robot does not need to be incapable of falling. It needs to be capable of continuing after it falls. That sounds obvious, but recovery occupies an awkward place in legged-robot research. Locomotion controllers are normally evaluated from valid standing states. Parkour policies may survive large impacts, but their training distributions still usually presume that the robot remains broadly controllable. Recovery begins where those assumptions break: the base may be inverted, several legs may be trapped beneath the body, the feet may have no meaningful support, and almost any link can be in contact with the environment. The problem becomes even harder away from a laboratory floor. On stairs, pipes, slopes, rocks, or irregular plates, “upright” is not a sufficient terminal condition. The robot must find a configuration that is dynamically quiet, mechanically feasible, and suitable for handing control back to a locomotion policy. A trunk aligned with gravity is not much use if one foot is suspended, a knee is wedged against an edge, or the robot immediately tips over again. Earlier terrain-aware recovery work, most notably DreamRiser, recognized this distinction. DreamRiser used proprioceptive history to infer latent terrain information and explicitly treated stable foot contact as part of recovery rather than optimizing base orientation alone. Its experiments demonstrated recovery on rough terrain, discrete obstacles, and slopes using Unitree A1 and Go1 platforms. Qian, Yin, and Kan attack the same broad problem but focus on a second weakness: even when reinforcement learning discovers a successful get-up motion, the resulting behavior may be unsuitable for a small physical robot. A policy can flail aggressively, exploit inaccurate collision models, or take an excessively cautious sequence that is technically successful but operationally slow. The paper’s central claim is that recovery needs both robustness and motion quality. Robustness comes from simulation, terrain inference, randomization, and curriculum learning. Motion quality comes from imitating demonstrations actually recorded on robot hardware. The result is a real-to-sim-to-real loop:. Observe viable recovery behavior in reality. Use that behavior as a prior while training a more general policy in simulation. Transfer the learned policy back to hardware without fine-tuning. The important point is that the real-to-sim phase is not merely system identification. Real data supplies a behavioral constraint: what a physically credible recovery should look like. Simulation then supplies the scale and diversity needed to turn that behavior into a robust skill. The Paper’s Main Contributions. The work combines four ideas that are individually familiar but unusually well matched to the recovery problem. First, it constructs an adversarial imitation objective from hardware-collected recovery demonstrations. Instead of learning every phase of self-righting from randomly initialized reinforcement learning, the policy receives a motion prior grounded in real actuator behavior. The authors argue that this produces more natural movement, improves learning efficiency, and reduces the energy cost associated with pathological simulated strategies. Second, it incorporates a terrain-imagination module. The controller estimates terrain-relevant latent properties from onboard measurements rather than assuming that the support surface is flat or relying on a conventional world-aligned terrain map. This matters because a fallen robot’s camera may be blocked, inverted, pointed at its own body, or pressed against the terrain. Third, the authors treat sim-to-real transfer as a contact-modeling problem as well as a dynamics-randomization problem. The framework combines domain randomization, a terrain curriculum, and what the paper calls geometry decomposition. In recovery, contacts occur over the trunk, hips, thighs, shanks, knees, and feet, so inaccurate link geometry can be substantially more damaging than it is during ordinary foot-dominated locomotion. Finally, the paper demonstrates direct deployment on a Unitree Go2, reports generalization across quadruped models, and presents real-world tests involving pipes, irregular plates, stairs, slopes, disturbances, and moving support. The authors position this as advanced all-terrain recovery on a comparatively small, affordable platform rather than a large custom machine with generous torque and thermal margins. The novelty is therefore not any single neural-network block. It is the division of labor among data sources: reality teaches the policy how to move, simulation teaches it where that motion must adapt, and onboard history tells it what kind of support it currently has. From Real Demonstrations to a Simulated Motion Prior. The first stage begins with recovery trajectories collected on physical hardware. The public materials do not specify whether those demonstrations come from Unitree’s stock recovery controller, a manually engineered controller, teleoperation, or a mixture. That detail will matter for exact reproduction, but the conceptual role of the data is clear. A hardware trajectory contains information that is hard to encode with a hand-designed scalar reward. It reflects real motor bandwidth, servo tracking, transmission compliance, current limits, latency, link collisions, and the timing of whole-body momentum transfer. Even if the demonstrator only works reliably on flat terrain, it provides a useful answer to a narrower question: among the many simulated ways to turn over, which motions resemble something a real robot can execute?. This is where adversarial imitation is a better fit than straightforward behavior cloning. Behavior cloning would ask the policy to reproduce demonstrated actions for demonstrated observations. That is problematic for recovery because the initial state distribution is enormous. A small difference in body orientation or contact state can make the demonstrated action inappropriate, and errors compound rapidly once the policy leaves the demonstration manifold. An adversarial motion objective instead learns a discriminator that distinguishes demonstration motion from policy-generated motion. The policy receives an additional reward when its state transitions resemble the hardware examples. It is not required to occupy the same state at the same timestamp or track a fixed trajectory frame by frame. It must reproduce the broader structure and style of the behavior while still solving the recovery task. That flexibility is critical. A fixed demonstration might roll the robot over its right shoulder, but an obstacle may block that route. Strict tracking would drive the robot into the obstacle. An adversarial prior can still favor coordinated, hardware-plausible movement while allowing the task objective to select a different contact sequence. The resulting policy should be understood as optimizing two coupled objectives. The task objective asks whether the robot reaches a stable recoverable configuration. The imitation objective regularizes how it gets there. Typical task terms in this class of controller reward uprightness, adequate body height, stable support, and a locomotion-ready joint configuration, while regularization discourages excessive torque, joint acceleration, action variation, and damaging impact. The adversarial reward replaces some of the brittle motion-shaping terms that would otherwise have to be tuned manually. This distinction explains the paper’s criticism of policies that are either aggressive or sluggish. A task-only reward often has broad equivalence classes. If success is defined as eventually becoming upright, violent momentum-driven strategies and slow quasi-static strategies may both score well. The real demonstration distribution narrows that equivalence class toward movements already shown to be executable on hardware. The same USTC group has previously used learned motion priors in quadruped loco-manipulation, combining trajectory-generated demonstrations with goal-conditioned reinforcement learning. Their recovery paper applies the general philosophy in a more contact-rich setting and, importantly, obtains the prior from the real robot rather than only from an optimizer or simulated reference. Terrain Imagination Without a Reliable World Map. A recovery controller has an unusual perception problem. While walking, a robot can point a depth camera or lidar toward upcoming terrain and maintain a gravity-aligned local map. After a fall, neither the viewing direction nor the visibility assumptions survive. The terrain immediately relevant to recovery may also be underneath the body, where cameras cannot see it. “Terrain imagination” addresses this by treating terrain as a hidden context inferred from temporal onboard observations. The phrase should not be read as requiring the reconstruction of a photorealistic or even metrically accurate height map. What the policy needs is a latent representation containing enough information to select an effective action. That representation might encode local surface orientation, the relative heights of supports, whether a limb is blocked, the presence of a nearby edge, effective friction, or the degree to which the trunk can roll in a particular direction. Contact itself becomes a sensor. If a commanded joint movement produces little motion but a large motor response, some part of the body is probably constrained. If one leg reaches support earlier than another, the local support heights differ. The sequence of joint motion, IMU readings, contact events, and previous actions can reveal properties that no single instantaneous observation makes explicit. The closest methodological ancestor is DreamWaQ and its recovery extension DreamRiser. DreamRiser used a context-aided estimator over proprioceptive history to predict body velocity and a latent terrain representation. During training, an asymmetric value function could see privileged simulator information, including local terrain measurements and applied disturbances, while the deployed policy operated on onboard signals. Qian and colleagues state that their terrain-imagination module predicts key terrain properties through onboard sensing. Although the public project page does not expose its tensor dimensions or exact supervision losses, the likely training logic is the same privileged-learning pattern: simulation provides labels or latent targets for terrain context, while the deployable estimator learns to infer that context from observation history. This module should improve recovery in two different phases. During the initial self-righting phase, it helps select a feasible roll or push direction. The robot may need to decide whether to press against the terrain with a knee, sweep a leg outward, or rotate the trunk around an existing contact. During the final stand-up phase, terrain context helps distribute the feet and choose a body pose compatible with the actual support plane. On a slope, aligning the trunk exactly as if the ground were flat may produce a marginal stance. On stairs or irregular plates, the nominal symmetric standing configuration may be physically impossible. The limitation is that this perception is inherently interactive. Before the robot moves or establishes informative contact, multiple terrain configurations may produce similar proprioceptive observations. Terrain imagination is therefore better described as learned contact-based state estimation than omniscient perception. It can become more accurate as recovery unfolds, but early actions still have to be robust to uncertainty. Why Geometry Matters More After a Fall. Ordinary locomotion simulation often concentrates fidelity on foot–ground contact and actuator dynamics. Recovery changes the dominant contact regime. A fallen quadruped can load the side of its trunk, an upper leg, a knee housing, or the edge of a shank. Small geometric differences change the moment arm around which the robot rolls. A simplified box may create a stable support face that does not exist on the real machine. Conversely, an overly conservative collision approximation may prevent a simulated leg from sliding through a configuration that is physically feasible. The paper lists geometry decomposition alongside domain randomization and curriculum learning as a transfer mechanism. The public page does not describe the precise decomposition algorithm, but its engineering role is clear: represent the robot’s complex contact surfaces in a form that the simulator can resolve accurately and efficiently. In practice, that generally means decomposing non-convex link meshes into simpler collision elements rather than representing an entire body link with one crude primitive. This is an important distinction because randomization cannot fix every modeling error. Varying friction, mass, motor strength, delay, and center of mass can make a policy tolerant to parametric uncertainty. It cannot compensate reliably for incorrect contact topology—such as a simulated flat face where the real robot has a rounded shell. The three transfer tools therefore address different failure modes:. Geometry decomposition improves the structure of body–terrain contact. Domain randomization covers uncertain physical and control parameters. Terrain curriculum learning makes optimization tractable as contact complexity grows. Curriculum learning is especially valuable here because dropping an untrained policy into arbitrary poses on severe terrain creates a difficult exploration problem. A sensible curriculum can begin with recoverable orientations and simple surfaces, then expand the distribution over initial attitudes, terrain relief, support discontinuities, friction, and disturbances. The policy first learns the semantic structure of “get upright and establish support” before being asked to solve that problem while wedged across stair edges. Domain randomization remains the final hedge against imperfect modeling. Its broader sim-to-real role is to force the learned policy to succeed over a family of plausible systems rather than over one nominal simulator. Training and Deployment. The authors report training in Isaac Gym using massively parallel simulation. The entire process takes approximately two hours on a single NVIDIA RTX 3090 and runs for 5,000 training iterations. That is a noteworthy engineering result: this is not a foundation-model-scale training exercise but a compact policy-learning pipeline that can plausibly be iterated by a robotics laboratory. The trained controller is then deployed directly on a Unitree Go2 without real-world fine-tuning. Sensor observations and policy actions are exchanged through ROS 2 and Unitree SDK2. The project also links deployment code, although it does not currently expose a complete end-to-end training implementation. The abstract additionally claims generalization to multiple quadruped robots. The confirmed physical deployment is the Go2; the public materials do not make equally strong hardware claims for every other model shown in the generalization demonstration. What the Experiments Actually Stress. The experiment suite is best understood as a matrix of contact conditions rather than a collection of visually impressive surfaces. The project first compares the proposed controller against two natural baselines: a predefined Unitree recovery trajectory and a conventional reinforcement-learning recovery controller. This comparison directly targets the paper’s motivating failure modes. The predefined sequence represents a reliable but environment-insensitive solution. The RL baseline represents an adaptive solution without the same hardware-derived motion prior. The proposed policy is intended to occupy the useful middle ground: adaptive without being violent, and agile without blindly replaying a trajectory. The authors then evaluate the policy on several real-world terrain categories. Pipe piles create curved and discontinuous supports. The robot may be balanced across multiple pipes, with contact normals that differ sharply from those of a flat floor. Rolling and sliding behavior becomes important, and small changes in link geometry can alter the recovery route. Irregular plates create piecewise-planar support with inconsistent heights and orientations. A fixed sequence cannot assume that the same knee, foot, or trunk contact will occur at the same time on every trial. Continuous stairs introduce edges, concave regions, and the possibility of a limb becoming wedged between levels. They also challenge the final transition into a locomotion-ready stance because the four feet may need to settle on multiple elevations. Grass slopes combine inclination with less predictable friction and compliance. This tests more than geometric adaptation: the policy must remain useful when its nominal contact model is wrong in several ways at once. The uncertainty tests add external disturbances and moving support. A disturbance during recovery changes angular momentum and invalidates open-loop timing. A moving base tests whether the controller is genuinely feedback-driven rather than merely executing a flexible version of a memorized sequence. The project also shows recovery performance in a QRC setting, providing an additional systems-level demonstration outside a carefully arranged flat-floor trial. Finally, the authors include a robot-generalization demonstration. This matters because a motion prior can easily overfit the proportions and actuator characteristics of the demonstration platform. Successful transfer to other quadruped models would suggest that the learned skill is represented at a more functional level—coordinated rolling, support creation, and standing—rather than as a sequence of platform-specific joint angles. Reading the Reported Results. The most concrete quantitative implementation result is the combination of short training time and zero-shot hardware deployment: roughly two hours of simulation training, followed by direct execution on the Go2 without real-world policy optimization. The primary performance result is breadth. The same controller is presented across substantially different contact geometries, disturbances, and support conditions. The paper reports that adversarial imitation promotes natural, energy-efficient motion, while terrain imagination and the transfer stack preserve robustness on complex surfaces. The authors further claim that the small Go2 demonstrates recovery behavior beyond that shown in prior work using more powerful hardware. That last statement should be read as the authors’ comparative claim rather than an independently standardized benchmark result. The baseline comparison is conceptually strong even without publicly visible aggregate numbers. A predefined trajectory should work when its assumed contact sequence is available and fail when terrain blocks that sequence. A task-only RL controller may adapt, but its movements can expose simulator exploitation, high action variation, or slow convergence to the final stance. The proposed method uses the demonstration prior to constrain the motion distribution while leaving enough freedom for terrain-dependent adaptation. The public project page does not provide numerical success counts, confidence intervals, recovery-time distributions, peak torque values, or energy totals. Those numbers are essential for deciding how large the practical gain is, especially when comparing “agility” against “hardware friendliness.” Consequently, the defensible conclusion is not that the method wins by a particular percentage. It is that the authors demonstrate a notably broad qualitative operating envelope, direct transfer, and a coherent causal explanation for the improvement. For future reporting, the most informative metrics would include recovery success within a fixed timeout, median and tail recovery time, mechanical work, peak motor current, impact impulse, maximum joint-speed violation, and the probability that a locomotion policy can take over without another fall. A recovery controller should be judged not only by whether it becomes upright, but by what it costs the hardware and what state it leaves behind. How This Fits Into the Literature. DreamRiser is the closest comparison on terrain awareness. It already established that proprioceptive terrain imagination can support all-terrain recovery. Qian and colleagues should therefore not be read as introducing terrain-aware recovery from nothing. Their advance is the combination of that idea with a real-hardware motion prior and a stronger contact-conscious sim-to-real pipeline. Adversarial motion priors provide the second lineage. Rather than specifying every aesthetic and coordination property with rewards, these methods learn a motion manifold from examples and use a discriminator-derived signal to keep the policy near that manifold. The authors’ earlier work on collision-free loco-manipulation similarly combines a learned motion prior with a task-directed controller, while their LEEPS work addresses autonomous terrain-conditioned parkour on inexpensive quadrupeds. The recovery paper brings these interests together: motion priors, environment-conditioned adaptation, and deployment on accessible hardware. What is distinctive is where the demonstration comes from. Animal motion capture can make a robot appear animal-like, and trajectory optimization can generate mechanically structured references. A real robot demonstration instead captures the behavior of the actual actuation and low-level control stack. The imitation data may be narrower, but it begins on the correct side of the reality gap. The method is therefore best viewed as robustification around a real behavioral seed. It does not ask simulation to invent the entire recovery strategy, and it does not ask a fixed real controller to handle every possible environment. Limitations and Practical Lessons. The first open question is demonstration provenance. The amount of data, number of initial poses, sensing used during collection, and identity of the demonstrating controller all affect reproducibility. A vendor get-up sequence provides a different prior from a human-designed optimization controller. It may also transfer vendor-specific biases into the learned policy. Second, adversarial rewards can produce their own shortcuts. A discriminator may focus on easily matched kinematic details rather than the physically important structure of the motion. For recovery, the discriminator input should ideally emphasize invariant relational features—link orientations, relative body geometry, contact transitions, and momentum patterns—rather than global position or an exact timing signature. Third, terrain imagination remains an estimate under partial observability. Some terrain properties are not identifiable until the robot probes the environment. Training should therefore reward safe information-gathering motions and tolerate uncertainty in the latent representation rather than forcing a single overconfident terrain reconstruction. Fourth, the confirmed real-world evidence centers on one small quadruped. Cross-model simulation is useful, but physical transfer across different transmissions, joint limits, shells, and torque envelopes would be substantially stronger evidence of morphological generalization. Finally, recovery is a safety-critical, contact-rich behavior. A policy may succeed repeatedly while accumulating damage through high-current events or shell impacts. Deployment needs independent current, temperature, joint-limit, and timeout guards. The recovery controller must also integrate with fall detection and locomotion handoff; otherwise, it remains a compelling isolated skill rather than a complete autonomy component. For teams adapting the approach, the practical recipe is straightforward:. Collect a small set of clean, physically credible recoveries on hardware. Represent those demonstrations with features that tolerate global-pose and timing differences. Train with both a task objective and an adversarial motion prior. Give the actor temporal onboard observations while allowing the critic or estimator privileged terrain information in simulation. Model collisions for every link likely to touch the world, not just the feet. Expand difficulty along separate curriculum axes: initial pose, terrain geometry, friction, disturbances, and actuator uncertainty. Evaluate hardware stress and post-recovery stability, not merely uprightness. The authors’ public repository currently offers project demonstrations and a deployment-code link, but not the complete training stack or visible quantitative tables. A fuller release would be particularly valuable because the difficult implementation details are likely to lie in collision geometry, demonstration preprocessing, termination conditions, and reward balancing rather than in the high-level network architecture. The Bottom Line. This paper’s strongest idea is not simply “use imitation” or “estimate the terrain.” It is the allocation of different kinds of knowledge to the domains that provide them most economically. Reality provides a compact prior over credible motion. Simulation provides the unsafe exploration and environmental diversity required for robustness. Terrain imagination converts partial onboard history into an actionable context. Geometry modeling, randomization, and curriculum then make the return trip to hardware plausible. For roboticists, the broader lesson is that recovery should not be treated as an emergency animation appended to a locomotion stack. It is a full-body, terrain-conditioned, contact-rich control problem—and one where a little carefully selected real data can constrain a great deal of simulated learning.