# Does Humanoid State Estimation Finally Have a Purpose-Built Solution?

A team of researchers led by Jixin Gao, Fucheng Liu, Teng Zhang, and Fusheng Zha has published KILVO — a state estimation framework that fuses all four sensor modalities already present on most commercial humanoid platforms (joint encoders, IMU, LiDAR, and camera) into a single asynchronous-sequential hybrid error-state iterated Kalman filter. The paper, posted to arXiv on August 7, 2026, reports competitive accuracy, efficiency, and output rates across multiple humanoid robots, gait patterns, and scenarios, with explicit resilience to sensor degradation and outright failures. Code and datasets are released on GitHub.

This matters because reliable odometry is the unglamorous prerequisite for everything else a walking robot must do. Whole-body control, loco-manipulation, and autonomous navigation all degrade catastrophically when state estimation drifts or fails — and humanoid platforms present challenges that standard mobile robot odometry frameworks were never designed to handle: dynamic leg occlusion, high-vibration IMU noise during bipedal impact, intermittent foot contact, and continuous body-height variation across different [gait cycle](https://humanoidintel.ai/glossary/gait-cycle) phases. KILVO is explicitly engineered around those constraints rather than adapted from wheeled-robot or quadruped solutions.

---

## What KILVO Actually Does — and Why the Architecture Choices Are Defensible

The core contribution is the asynchronous-sequential hybrid ESIKF design. Here is how the data flow works according to the paper:

1. **IMU data drives prediction** — the highest-frequency signal propagates the state forward in time continuously.
2. **Leg kinematics are processed asynchronously at a high rate**, providing [proprioception](https://humanoidintel.ai/glossary/proprioception)-based constraints without stalling the filter waiting for slower exteroceptive updates.
3. **LiDAR updates are applied first** among exteroceptive modalities, registering point clouds for geometric priors.
4. **Camera updates follow sequentially**, correcting the state via photometric errors after LiDAR has already tightened the geometric estimate.

The ordering is deliberate. LiDAR geometry is generally more pose-observable than raw photometric data in a single update step, so using it first conditions the visual update on a tighter prior — reducing the risk of the visual component latching onto a local minimum in the photometric cost surface.

A compact contact estimation module is embedded within the framework. Critically, the authors note it shares information with state estimation without requiring additional sensors beyond those already enumerated. That is non-trivial: on a legged system, knowing which feet are in contact — and how confidently — directly affects how much weight to give kinematic odometry versus inertial propagation. Doing this without a dedicated force/torque sensor or contact microphone array keeps the system deployable on hardware that doesn't have those instruments.

The multimodal adaptation capability deserves particular attention. The paper describes the framework as resilient to sensor failures, meaning individual modality dropouts do not cause full odometry collapse. For operators deploying humanoids in unstructured environments — dusty warehouses where LiDAR returns degrade, or visually featureless corridors where cameras provide weak constraints — this is an operationally significant property. Most published humanoid state estimators are fragile in exactly these conditions.

---

## The Validation Picture — and Where to Be Skeptical

The paper validates KILVO on public datasets and in real-world experiments across multiple humanoid robots and gait patterns. The authors claim it outperforms "state-of-the-art fusion methods" in accuracy, efficiency, and output rates under sensor degradation conditions.

A few healthy caveats apply before reading too much into the comparison:

**Which humanoid robots?** The abstract does not name specific platforms. The broader humanoid field spans dramatically different kinematic structures — a platform like [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics)' H1/G1 series has different vibration profiles and sensor placement constraints than a taller, heavier industrial humanoid. Cross-platform generalization claims need platform names attached to be fully credible.

**What counts as "competitive" accuracy?** Without trajectory RMSE numbers grounded in the source abstract, we cannot make quantitative comparisons here. The full paper presumably contains these — readers evaluating KILVO against alternatives like VILENS, DILIGENT-KIO, or other legged odometry frameworks should go directly to the paper's experimental section for the numbers.

**Real-world test conditions?** "Multiple scenarios" covers a lot of ground. Whether these include the kind of aggressive dynamic motions — fast lateral stepping, stair climbing, recovery from stumbles — that stress odometry hardest is something the full paper must be evaluated on.

The GitHub code release is the most concrete signal of the team's confidence. Open-sourcing a state estimation framework invites independent replication, which is a significantly stronger credibility marker than a closed-source benchmark.

---

## Why This Matters for the Broader Industry

State estimation for humanoids has been an underinvested area relative to the attention paid to locomotion policies, manipulation learning, and perception. The dominant research energy has gone into sim-to-real transfer for control policies, VLA architectures, and whole-body control formulations. But a locomotion policy running on a robot that doesn't know where it is — or that loses track of its pose during a gait disturbance — fails in deployment even if the policy itself is excellent.

The practical consequence: as companies including [Agility Robotics](https://humanoidintel.ai/companies/agility-robotics), [Figure AI](https://humanoidintel.ai/companies/figure-ai), and [Boston Dynamics](https://humanoidintel.ai/companies/boston-dynamics) push humanoids into real operational environments, the state estimation layer becomes a critical failure mode. KILVO's multimodal fusion approach — particularly the contact estimation module that avoids additional hardware — is exactly the kind of infrastructure-level contribution that deployment teams need but rarely get from academic robotics.

The code release on GitHub means engineering teams can begin integration and stress-testing immediately. If the real-world performance holds up under independent evaluation, KILVO has a realistic path to becoming a community baseline for humanoid odometry in the same way that established frameworks became defaults in the mobile robot space.

---

## Key Takeaways

- **KILVO** is a new state estimation framework from Gao, Liu, Zhang, and Zha, designed specifically for humanoid robots fusing joint encoders, IMU, LiDAR, and camera.
- The architecture uses an **asynchronous-sequential hybrid ESIKF**: IMU predicts, kinematics update asynchronously, then LiDAR and camera update sequentially in that order.
- A built-in **contact estimation module** integrates with state estimation without adding hardware sensors.
- **Multimodal adaptation** allows the filter to degrade gracefully rather than fail catastrophically when individual sensors malfunction.
- Validated across **multiple humanoid robots, gait patterns, and real-world scenarios**; code and datasets are publicly released on GitHub.
- The absence of named platform specifics and numeric accuracy benchmarks in the abstract means independent validation of the full paper's claims is warranted before treating the performance comparisons as definitive.

---

## Frequently Asked Questions

**What is KILVO?**
KILVO is a kinematic-inertial-LiDAR-visual odometry system for humanoid robots, developed by Jixin Gao, Fucheng Liu, Teng Zhang, and Fusheng Zha. It fuses joint encoder data, IMU, LiDAR, and camera inputs within an asynchronous-sequential hybrid error-state iterated Kalman filter to estimate a humanoid robot's pose and motion with high accuracy and resilience to sensor failures.

**How does KILVO differ from standard robot odometry frameworks?**
Unlike odometry systems designed for wheeled robots or quadrupeds, KILVO is purpose-built for bipedal humanoids. It accounts for humanoid-specific challenges including dynamic leg occlusion, intermittent foot contact, and IMU noise from bipedal impact, and it incorporates a contact estimation module that does not require additional force/torque sensors.

**Why does KILVO update LiDAR before camera in the filter?**
The sequential update order — LiDAR first, then camera — is deliberate. LiDAR point cloud registration provides geometric constraints that tighten the pose estimate before the photometric (visual) update is applied, reducing the risk of the visual component converging to incorrect local minima.

**Is KILVO's code publicly available?**
Yes. According to the paper, code and datasets are released on GitHub.

**What is the significance of multimodal adaptation in humanoid state estimation?**
Multimodal adaptation means KILVO can continue producing useful pose estimates even when individual sensors degrade or fail entirely — for example, in dusty environments where LiDAR returns deteriorate, or in visually featureless spaces where camera constraints are weak. This is operationally critical for humanoids deployed in unstructured real-world environments.