Two New Papers Push Quadrotor RL Further Into the Real World. Here's What They Actually Show.

The authors' solution is a recurrent policy trained within an asymmetric actor-critic architecture. The asymmetry is important here: the critic has access to privileged state information during training that the actor does not have at deployment. This is a well-established technique in the sim-to-real literature, appearing in work from Ashish Kumar and colleagues at Berkeley and in various legged locomotion papers from ETH Zurich, but it is genuinely appropriate for this setting given the severity of the partial observability problem.

What I find more interesting is the decision to use an Incremental Nonlinear Dynamic Inversion (INDI) controller as an inner loop to track the policy's output, rather than mapping directly from policy to motor commands. INDI is an incremental variant of NDI that estimates and cancels model uncertainty online, and it has been gaining traction in agile drone control research over the past several years. The combination of a learned outer loop with a model-based inner loop is not new as a concept, but the specific implementation here, where INDI is used to close the gap between a recurrent RL policy and real hardware, is a reasonable engineering choice.

The zero-shot transfer results are the headline claim, and they do hold up under scrutiny, with the caveat that the experimental conditions, while varied (different initial attitudes, added payloads, wind disturbances), are still lab conditions. It's worth noting that the paper does not test on particularly extreme wind or on surfaces other than presumably flat ground. Whether this generalises to, say, a rocky surface or a sloped landing area remains unclear.

What the aerial manipulation paper is doing differently

The second paper, arXiv:2606.16621, tackles a harder problem in some respects. Aerial manipulation, meaning a flying robot that can physically interact with its environment using an attached arm, has been an active research area for over a decade. The core difficulty is that the arm and the vehicle are dynamically coupled: moving the arm changes the vehicle's inertia and center of mass, and the vehicle's motion disturbs the arm. Most prior work either simplifies this coupling or relies on slow, conservative arm motions to avoid it.

This paper takes a different approach. The RL outer loop maps desired end-effector targets in 6-DOF directly to whole-body commands, meaning it is learning to coordinate arm and vehicle motion together rather than treating them as separate problems. The inner loop then handles the low-level compensation using a dynamics estimator that does not require a pre-specified system model. The estimator is essentially doing online identification of the transient inertial shifts caused by rapid arm motion and payload changes.

The baseline comparisons are where this paper earns some credibility. The authors compare against RL with a PID inner loop and against RL with an INDI-plus-PID inner loop. The proposed method reduces end-effector tracking error and improves task success rate across the tested hardware conditions. I know I'm being picky here, but the paper does not specify exactly how many trials constitute "the tested hardware conditions," and the hardware is a single custom quadrotor platform with a 3-DOF manipulator. The sample size is small, and this has not been replicated on other platforms.

That said, the result that a dynamics estimator outperforms both PID and INDI baselines in this setting is interesting. INDI is generally considered a strong baseline for aerial vehicles because of its model-free disturbance rejection properties, so beating it, even on a limited hardware setup, suggests the estimator is doing something genuinely useful.

The sim-to-real question

Both papers are dealing with the same fundamental challenge in different ways, namely the gap between simulation and reality. This is, actually, the research shows, still one of the most important open problems in learned robot control, and the strategies being used here reflect a broader trend in the field.

The fall recovery paper invests heavily in simulation fidelity. The authors describe high-fidelity models of motor response and optical flow, and they run ablation studies to validate each design choice. This is the right methodology. Ablation studies in RL papers are sometimes perfunctory, included to satisfy reviewers rather than to genuinely test the contribution, but the ones described here seem to address the actual design decisions that matter: the recurrent architecture, the INDI inner loop, the asymmetric critic.

The aerial manipulation paper takes a somewhat different approach, leaning more on the inner-loop estimator to handle the residual sim-to-real gap rather than trying to close it entirely in simulation. This is a pragmatic choice, and it arguably makes the system more robust to model errors that are difficult to simulate, such as the exact coupling dynamics of a specific arm-vehicle combination. The tradeoff is that the estimator itself needs to converge quickly enough during execution to be useful, and the paper does not provide a detailed analysis of estimator convergence time under different conditions.

What is genuinely new versus what is incremental

This is a distinction I think is worth making explicitly, because it affects how you should interpret both papers.

The aerial manipulation paper is, to be honest, incremental over prior work on hierarchical RL-plus-model-based control for aerial robots. The combination of a learned outer loop with an estimator-based inner loop is a sensible and well-motivated contribution, and the hardware results are useful, but the core architectural idea has precedents in the literature on adaptive control and in prior aerial manipulation work from groups including those at LAAS-CNRS and ETH's Autonomous Systems Lab. The novelty is in the specific combination and in the hardware validation.

The fall recovery paper is, I think, more genuinely novel in its scope. Autonomous fall recovery for quadrotors from arbitrary ground attitudes is not a well-studied problem. Most prior work on quadrotor recovery assumes the vehicle is already airborne or addresses specific, limited failure modes. The combination of recurrent policy, asymmetric training, INDI inner loop, and zero-shot transfer for this specific problem represents a meaningful advance, even if each individual component has been used elsewhere.

It's worth noting that bidirectional thrust, mentioned in the fall recovery paper's title, refers to the ability of the motors to spin in both directions, which expands the recovery maneuver space significantly. This is a hardware capability that is not universal, and results may not transfer to standard unidirectional-thrust platforms without modification.

Why these papers matter together

Reading these two papers together reveals something about the current state of RL for aerial robotics that is worth stating plainly. The field has moved past the phase where the interesting question was simply whether RL policies could transfer to real hardware at all. That question has been answered affirmatively, repeatedly, across legged robots, manipulation arms, and aerial vehicles. The interesting questions now are more specific: which combinations of learned and classical control work best for which problem structures? How much simulation fidelity is enough? What are the failure modes of recurrent policies under sensor degradation?

Both papers are engaging with these questions seriously. Neither is claiming to have solved aerial robotics. The fall recovery paper is solving a specific, well-defined problem with careful engineering. The aerial manipulation paper is making a targeted improvement over existing baselines in a domain where progress has been slow. That is basically what good research looks like.

The broader implication, and this is speculative, is that the most productive direction for RL in aerial robotics may not be end-to-end learned control but rather hybrid architectures where the learned component handles high-level coordination and the classical component handles low-level robustness. Both papers support this view, from different angles.

Open questions and what I would want to see next

Several things remain unclear from both papers, and they are worth naming.

For the fall recovery work: how does performance degrade as the initial attitude becomes more extreme? The paper tests a range of attitudes, but it is not obvious from the abstract whether truly inverted configurations (fully upside-down) are included. The INDI inner loop's performance under severe motor asymmetry, which might occur if one motor is damaged during the fall that caused the recovery scenario in the first place, is also not addressed. That is arguably the most practically relevant failure mode.

For the aerial manipulation work: the 3-DOF arm is a relatively simple manipulator. It's too early to say whether the dynamics estimator approach will scale to higher-DOF arms where the coupling dynamics are more complex and the estimator's task becomes correspondingly harder. The paper also does not address what happens when the payload changes mid-task, as opposed to between trials, which is a scenario that would occur in real-world pick-and-place operations.

More broadly, both papers would benefit from longer-horizon deployment testing. Lab experiments over tens or hundreds of trials are necessary but not sufficient for understanding how these systems behave over thousands of cycles, where wear, calibration drift, and environmental variation start to matter.

I would also want to see both approaches tested on platforms other than the specific hardware used in each paper. The fall recovery work uses a bidirectional-thrust quadrotor, which is a somewhat specialised platform. The manipulation work uses a custom vehicle. Replication on commercial or widely-available platforms would significantly strengthen the claims about generalisability.

None of this is a criticism of the papers as papers. Both appear to be doing what they set out to do, with appropriate scope and honest methodology. The limitations I have described are the natural next steps, not failures of the current work. That distinction matters, and it tends to get lost when research coverage focuses on the headline result rather than the research trajectory.

Aisha Patel · 11 hours ago · 10 min