# Does Online Parameter Estimation Finally Solve Legged Loco-Manipulation's Sim-to-Real Problem?

A new framework from researchers Hossein Keshavarz, Alejandro Ramirez-Serrano, and Majid Khadiv matches the performance of a controller with access to ground-truth physical parameters — without ever knowing those parameters in advance. That is the central claim of **Adaptive-MHE**, a paper posted to arXiv on September 17, 2026, which proposes an online, sampling-based system identification framework built on moving horizon estimation (MHE) and coupled with a sampling-based model predictive controller.

The core problem it targets is one of the most stubborn in humanoid and legged robotics: when a robot picks up or pushes an object, or steps onto unfamiliar terrain, it typically has no reliable knowledge of that object's mass or the surface's friction coefficient. Mismatches between these real-world parameters and the simulated counterparts used to design the controller produce a [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) gap that degrades [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation) performance — sometimes catastrophically. Adaptive-MHE estimates both object and terrain parameters online, continuously, as conditions change — and the authors report that in both simulation and hardware experiments, the framework consistently outperforms baseline approaches.

---

## Why Classical System Identification Breaks Down for Contact-Rich Robots

Traditional system identification methods carry a foundational assumption: that the dynamics of the system are differentiable. This is a reasonable approximation for robotic arms operating in free space, but it does not hold for legged robots. Contacts with the ground, objects, and obstacles introduce discontinuities — the equations of motion are non-smooth at contact transitions, and gradient-based Sys-ID methods simply cannot navigate that terrain reliably.

Sampling-based Sys-ID sidesteps this by skipping gradients entirely. Instead of differentiating through dynamics, it runs massively parallel simulation rollouts and directly matches simulated state trajectories against recorded real-world trajectories. The parameter set that produces the best match wins. This approach has clear advantages for contact-rich legged systems where differentiability fails.

The limitation, as Keshavarz et al. note, is that existing sampling-based Sys-ID approaches have typically been deployed **offline** — you run the identification procedure before deployment, lock in parameters, and hope the real world doesn't change. For a humanoid carrying objects of varying weight across different floor materials, that assumption breaks immediately.

---

## What Adaptive-MHE Actually Does

The key architectural contribution is coupling online Sys-ID with a moving horizon estimation framework. Rather than treating parameter estimation as a one-shot offline step, MHE maintains a sliding window of recent observations and continuously re-solves for the most consistent parameter estimates given that history. As conditions change — the robot picks up a heavier object, transitions from carpet to tile — the estimates update in real time, and the sampling-based MPC controller receives fresh parameters on each planning cycle.

The physical parameters being estimated include, per the paper, quantities like **mass and friction** — the two most consequential unknowns for [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) of legged manipulators. Estimating friction online is particularly significant: friction coefficients are notoriously difficult to measure ahead of time and vary continuously with surface condition, load, and contact geometry.

The framework is positioned explicitly against the [sim-to-real](https://humanoidintel.ai/glossary/sim-to-real-transfer) gap that afflicts both classical controllers and learned policies. Where a VLA or behavior-cloning approach might rely on [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) across unseen parameters, Adaptive-MHE takes the explicit estimation route — measure the gap and close it numerically, rather than hoping the policy generalizes.

---

## Hardware Validation: What the Paper Claims and What It Doesn't

The authors report that Adaptive-MHE was tested in both simulation and hardware experiments, with the framework "consistently outperforming baselines and matching the performance of a controller with access to ground-truth parameters." That is a meaningful benchmark — ground-truth access is the theoretical ceiling for any parameter-dependent controller.

**A note of caution is warranted here.** The paper is a preprint posted to arXiv and has not yet undergone peer review. The hardware platform used, the specific legged robot system tested, the nature of the manipulation tasks, and the quantitative margin of improvement over baselines are not specified in the abstract. Before these results can be treated as fully validated, the robotics community will need to see the full experimental methodology: which robot, which tasks, what baseline controllers, and whether the "matching ground-truth performance" claim holds across a representative range of parameter variation rather than a narrow experimental regime.

This is not a dismissal — the approach is technically coherent and the MHE formulation is principled — but single-paper hardware claims in legged locomotion research have a mixed track record of generalizing beyond the specific lab conditions in which they were generated.

---

## Why This Matters for the Humanoid Industry

The [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation) problem sits at the intersection of two capabilities every humanoid platform is racing to demonstrate at deployment scale. Walking is largely solved at the research level; manipulating objects while walking in unstructured environments is not. The parameter uncertainty problem — not knowing the mass of what you're carrying or the friction of what you're walking on — is a concrete, daily-use failure mode, not a theoretical edge case.

Current commercial approaches to this problem tend to fall into two camps: (1) heavily engineered perception pipelines that attempt to estimate object properties from vision before contact, and (2) large learned policies trained across massive parameter distributions in hopes of implicit generalization. Adaptive-MHE proposes a third path: explicit online estimation tightly integrated with the planning loop. If the hardware results generalize, this approach has practical relevance for any legged platform doing real-world object interaction — from warehouse floor operations to household assistance.

The sampling-based MPC backbone is also worth noting. Frameworks like MPPI (Model Predictive Path Integral) have gained significant traction in legged robotics precisely because they handle non-smooth dynamics gracefully. Building adaptive Sys-ID on the same sampling foundation is architecturally clean and avoids introducing a differentiability requirement through the back door.

For engineering teams at companies pushing [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) on humanoid platforms, this paper is worth a close read regardless of peer review status — the MHE coupling mechanism and the online sampling Sys-ID formulation represent a concrete design pattern applicable to any legged manipulator facing unknown environmental parameters.

---

## Key Takeaways

- **Adaptive-MHE** is an online, sampling-based system identification framework that estimates physical parameters (mass, friction) in real time and feeds them to a sampling-based MPC controller for legged loco-manipulation.
- The framework is designed specifically to handle **contact-rich legged dynamics**, where classical differentiable Sys-ID methods fail.
- In simulation and hardware experiments, it **matches the performance of a controller with access to ground-truth parameters** — a result that, if it holds under peer review scrutiny, is significant.
- Unlike offline Sys-ID, Adaptive-MHE **adapts continuously** as environmental conditions change during a task.
- The paper is an **arXiv preprint** (September 17, 2026); hardware platform, task specifics, and quantitative margins over baselines are not detailed in the abstract and require the full paper for complete evaluation.
- The approach is directly relevant to any commercial humanoid platform pursuing real-world loco-manipulation where object and terrain properties are unknown at runtime.

---

## Frequently Asked Questions

**What is Adaptive-MHE and what problem does it solve?**
Adaptive-MHE is an online parameter estimation framework for legged robots. It estimates physical properties like mass and friction in real time using moving horizon estimation, then feeds those estimates to a model predictive controller. This closes the sim-to-real gap that occurs when a robot's controller was designed with inaccurate assumptions about the environment.

**Why can't standard system identification methods work for legged robots?**
Classical Sys-ID assumes differentiable dynamics, but legged robots in contact with the ground and objects have non-smooth, discontinuous dynamics at contact transitions. Sampling-based methods avoid this by matching simulated and real trajectories directly through parallel rollouts, without needing gradients.

**What does "moving horizon estimation" mean in this context?**
MHE maintains a sliding window of recent sensor observations and continuously re-solves for the physical parameters most consistent with that recent history. This allows the system to track changing conditions — such as picking up a heavier object or stepping onto a different surface material — rather than committing to a fixed parameter estimate set offline.

**How does Adaptive-MHE differ from learning-based approaches to sim-to-real transfer?**
Learning-based approaches (VLAs, reinforcement learning with domain randomization) aim for implicit generalization across parameter distributions. Adaptive-MHE takes an explicit estimation approach: it actively measures the current parameter mismatch and corrects for it numerically each planning cycle. The two approaches are not mutually exclusive and could be combined.

**Has this been validated on real hardware?**
The authors report hardware experiments in which the framework consistently outperforms baselines and matches ground-truth-parameter controller performance. However, the paper is a preprint and has not yet been peer reviewed. The specific hardware platform and full experimental details are available in the complete paper at arxiv.org/abs/2609.17832.