UMI-on-Air: Embodiment-Aware Aerial Systems

• UMI-on-Air is an integrated framework that leverages generalizable, embodiment-agnostic policies to adapt robotic manipulation and wireless relay tasks across diverse aerial platforms.
• It employs a two-tier system with a high-level diffusion-based policy for trajectory planning and a low-level embodiment-specific controller for real-time execution.
• In wireless communications, the framework optimizes UAV-mounted IRS configurations by balancing beam directivity with robustness against platform fluctuations.

UMI-on-Air refers to two key frameworks developed for advancing aerial robotics and wireless communications: 1) a method for embodiment-aware deployment of embodiment-agnostic visuomotor manipulation policies for robots (notably unmanned aerial manipulators, UAMs), and 2) a framework for optimizing UAV-mounted intelligent reflecting surface (IRS) systems for robust wireless relaying under hovering fluctuations. Both research threads share the central theme of leveraging platform-independent policies or hardware to achieve high-performance, generalizable behaviors in highly dynamic aerial environments (Gupta et al., 2 Oct 2025, Zakavi et al., 23 Apr 2025).

1. Universal Manipulation Interface (UMI) and the Embodiment Gap

Universal Manipulation Interface (UMI) policies are trained on diverse, unconstrained human demonstrations using a lightweight hand-held gripper equipped with egocentric vision and SLAM-based pose tracking. These demonstrations are collected in-the-wild and are not biased toward any specific robotic embodiment or platform. A critical challenge emerges when deploying these policies on robotic platforms with substantial actuation constraints, limited dynamics (e.g., UAMs with underactuation, aerodynamic effects, and control bounds), or in new environments. The core problem, termed the "embodiment gap," arises from this mismatch between the generality of the trained policy and the highly-specific capabilities and constraints of the deployment machine. The goal is to close this gap at inference time—enabling high-level policies to adapt seamlessly and safely to new robotic platforms—without retraining or additional data collection (Gupta et al., 2 Oct 2025).

2. System Architecture: High-Level Policy and Low-Level Controller

UMI-on-Air decomposes the robotic system into a two-level hierarchy:

A. High-Level UMI Policy (Embodiment-Agnostic):

• Trained offline using behavior cloning on egocentric RGB observations and gripper kinematic data, collected with the UMI device.
• Policy architecture employs a conditional UNet-based diffusion model to generate trajectory sequences $a = \{p_{r_t}, R_{r_t}, w_{r_t}\}_{t=1}^W$, representing desired end-effector positions ($p_r$), orientations ($R_r$), and gripper widths ($w_r$).
• Multimodality emerges because the policy is uninformed by embodiment constraints during learning and may propose infeasible actions for a given robot.

B. Low-Level, Embodiment-Specific Controller:

• Receives a reference trajectory $a$ and executes it in real time.
• Instantiation options include inverse kinematics (IK) with velocity limits for fixed-base arms or model-predictive control (MPC) for UAMs.
• Tracking cost $L_{track}(a)$ quantifies the difficulty of following trajectory $a$ given embodiment dynamics.
• MPC-based controllers minimize quadratic costs in end-effector tracking and control, respecting robot-specific constraints at high frequencies (up to 50 Hz) (Gupta et al., 2 Oct 2025).

3. Embodiment-Aware Diffusion Policy (EADP): Algorithmic Coupling

The Embodiment-Aware Diffusion Policy (EADP) framework is a closed-loop inference mechanism in which the low-level controller provides online feedback to the diffusion policy, biasing trajectory sampling toward dynamically-feasible regions.

• Diffusion Sampling: Employs Denoising Diffusion Implicit Models (DDIM). At each step, the policy denoiser $\pi_\theta$ predicts a cleaner trajectory from noisy samples.
• Gradient Feedback: At each diffusion iteration $k$, the gradient $\nabla_{a^k} L_{track}(a^k)$ of the controller's tracking cost is computed.
• Guidance Step: A classifier-like update nudges the sample:

$$\tilde{a}^k = a^k - \lambda \bar{\omega}_k \nabla_{a^k} L_{track}(a^k)$$

where $\lambda$ is the guidance scale and $\bar{\omega}_k$ reflects the diffusion noise schedule.

• Two-stage Update: After guidance, the denoiser is applied; this process is iterated from high to low noise, with guidance strength increasing as trajectory refinement progresses.
• Algorithmic Outline:

def EADP_Sample(o, λ): a[K] ← sample Normal(0, I) for k = K down to 1: cost ← ControllerTrackingCost(a[k]) g ← ∇_{a[k]} cost ã[k] ← a[k] − λ * ω[k] * g a[k−1] ← ã[k] + DDIM_Update(π_θ(ã[k], k | o)) return a[0]

(Gupta et al., 2 Oct 2025)

This approach enables plug-and-play, embodiment-aware trajectory adaptation without additional data or retraining.

4. Empirical Evaluation: Manipulation and Communication Tasks

A. Embodiment-Aware Policy Transfer for Manipulation

• Simulation tasks (Open-and-Retrieve, Peg-In-Hole, Rotate-Valve, Pick-and-Place) revealed that naïve diffusion policies underperform when embodiment constraints and disturbances (e.g., UAM base noise) are present.
• EADP yields consistent improvements in task success rates across UR10e arms and UAMs, with robustness to disturbances and transfer to unseen domains.
• Gains are substantial especially for aerial platforms under disturbance (average +20% success rate increase versus unguided baselines).

B. UAV-Mounted IRS Performance

• In the communications domain, UMI-on-Air models an IRS mounted on a hovering UAV acting as a relay between base station (BS) and user equipment (UE).
• Performance is dominated by the UAV’s angular vibrations (modeled as Gaussian $\epsilon_x$, $\epsilon_y$).
• End-to-end IRS gain $G$ is characterized by a closed-form mixture of perturbed main lobe and side-lobe contributions, derived from fluctuations in UAV pose.
• Outage probability $P_{out}$ is analyzed under passive and active IRS scenarios, using CLT and Gamma approximations for the composite channel coefficients (Zakavi et al., 23 Apr 2025).

Experimental Setting	Naïve Diffusion Success (%)	EADP Success (%)
UAM + Disturbance (Sim.)	40–58	65–80
Peg-In-Hole (Aerial, Real)	0	100
Lemon Harvest (Aerial)	20	80

Table: Representative success rates from (Gupta et al., 2 Oct 2025); values per manipulation task and embodiment.

5. Communication System Modeling and Outage Analysis

UMI-on-Air's communication channel analysis involves:

• 3D Pattern Perturbation: Derivation of the normalized IRS gain $G$ as a function of angular fluctuations, decomposed into sectorized mixture distributions for tractable analysis.
• Outage Probability Expressions: For SISO-Passive-IRS,
  • CLT: $$P_{out} \approx \sum p_n Q\left( \frac{\mu_v - \sqrt{\gamma_{th}/(\gamma_0 q_n)}}{\sigma_v} \right)$$
  • Gamma: $$P_{out} = \sum p_n \left[ \gamma\left( \Lambda, \sqrt{\gamma_{th}/(\gamma_0 q_n)}/\Omega \right) / \Gamma(\Lambda)\right]$$
• Passive vs. Active IRS Elements: More elements increase peak gain but decrease beamwidth, thus enhancing sensitivity to UAV oscillations. Active elements compensate path loss but introduce amplification noise.

Key findings indicate that, under UAV fluctuations, increasing the number of IRS elements is not always beneficial; an optimal $N^*$ exists that balances directivity and outage performance. The optimal number of addressed elements decreases as fluctuation variance grows (Zakavi et al., 23 Apr 2025).

6. Design Guidelines for Aerial Deployment

For robotic manipulation:

• No additional data collection or retraining is required to adapt UMI policies to new robotic embodiments.
• Embodiment-aware guidance enables robust deployment in unseen environments.
• Practical implementations recommend on-board vision systems matched to the demonstration setup, SLAM for pose estimation, and fast trajectory-tracking controllers.

For wireless communications:

• Characterize UAV angular stability and sectorize IRS element patterns according to fluctuation levels.
• Use closed-form mixture or Gaussian-based outage formulas to select the optimal number of IRS elements.
• Under severe jitter, favor fewer IRS elements with wider beams, complemented by active components if power permits.
• Dynamically adapt the number of enabled elements in real time as flight stability changes (“on-air” adaptation).

7. Outlook and Practical Impact

UMI-on-Air establishes a generalizable methodology for both manipulation and communication systems that decouples high-level policy learning or beamforming from embodiment/performance constraints, closing the gap at deployment via two-way feedback or probabilistic pattern modeling. In manipulation, this approach substantially improves aerial robot capability in the face of strong embodiment constraints and disturbances. In communications, it enables reliable relay links by balancing beam directivity with robustness to platform fluctuations. The plug-and-play nature and avoidance of retraining suggest broad applicability across unseen embodiments and dynamic environments (Gupta et al., 2 Oct 2025, Zakavi et al., 23 Apr 2025).