UDON: Uncertainty-weighted Distributed Optimization for Multi-Robot Neural Implicit Mapping under Extreme Communication Constraints

1. Problem Statement

Multi-robot systems utilizing Neural Implicit Representations (e.g., Neural Radiance Fields or NeRFs) offer compact and high-fidelity 3D mapping capabilities. However, these collaborative systems are highly vulnerable to severe communication constraints, such as limited bandwidth, high latency, and packet loss.

• The Challenge: Standard distributed optimization methods (like Consensus Alternating Direction Method of Multipliers, or C-ADMM) rely on frequent information exchange to reach a consensus. Under extreme conditions (e.g., communication success rates below 5% or even 1%), these methods often fail to converge, leading to divergent maps, reconstruction artifacts, or incomplete scene representations.
• Limitations of Prior Work:
  • Methods like RAMEN use uncertainty-weighted consensus but degrade significantly below 20% success rates.
  • Methods like MACIM and OpenMulti use consistency losses to handle failures but often lack the ability to capture fine environmental details.
  • Existing approaches often use a single aggregate dual variable per agent, causing outdated information from disconnected neighbors to destabilize the optimization process.

2. Methodology: UDON Framework

The authors propose UDON, a real-time multi-agent framework that integrates uncertainty-weighting with a novel edge-based distributed optimization strategy.

A. Neural Implicit Representation

UDON builds upon the Co-SLAM framework. Each agent maintains a local map parameterized by:

• A multi-resolution hash grid ($V_{\Theta_i}$) for feature storage.
• Pre-trained geometry and color MLP decoders (fixed parameters).
• Only the hash grid parameters are updated during mapping.
• Rendering is performed via volumetric ray casting.

B. Frequency-Based Uncertainty Modeling

To determine which parts of a map are reliable, UDON employs a frequency-based uncertainty model:

• Concept: The number of gradient updates a specific hash grid vertex receives serves as a proxy for its certainty. Vertices with many updates (observed frequently) have low uncertainty; those with zero updates have high uncertainty.
• Mechanism: Agents compute an uncertainty vector ($u_i$) tracking non-zero gradient counts. When two agents communicate, they generate diagonal weight matrices ($W_{ij}, W_{ji}$) based on normalized uncertainty. This ensures that during fusion, reliable (low uncertainty) map regions are trusted more than unreliable ones, preventing fine details from being "averaged out."

C. Uncertainty-Weighted Distributed Optimization

This is the core innovation of UDON, addressing the instability of standard C-ADMM under packet loss.

1. Edge-Specific Dual Variables: Unlike prior methods that use a single dual variable ($p_i$) per agent (accumulating errors from all neighbors), UDON assigns a unique dual variable $p_{(i,j)}$ to each communication edge $(i, j)$.
2. Selective Updates:
   • If agents $i$ and $j$ are communicating, $p_{(i,j)}$ is updated using the uncertainty-weighted consensus rule.
   • If they are not communicating (packet loss), $p_{(i,j)}$ remains unchanged.
3. Primal Update: The local map update minimizes a loss function combining:
   • Local reconstruction error.
   • Consensus terms weighted by the current dual variables of active neighbors only.
   • Benefit: This prevents the optimization from being penalized by outdated information from long-disconnected neighbors, a common cause of divergence in sparse communication.

3. Key Contributions

1. UDON Framework: A novel uncertainty-weighted distributed optimization method specifically designed for extreme communication scenarios (down to 1% success rate).
2. Edge-Based Dual Variables: A departure from aggregate dual variables, assigning independent dual variables to communication edges to isolate and penalize disagreement only between currently connected pairs, ensuring stability.
3. Empirical Validation: Extensive experiments on synthetic benchmarks (Replica, ScanNet) and real-world TurtleBot hardware, demonstrating superior performance over state-of-the-art baselines (Di-NeRF, RAMEN, MACIM).

4. Experimental Results

The authors evaluated UDON across varying communication success rates (1% to 20%) and agent counts (4 to 8).

• Extreme Communication (1% Success Rate):
  • Baselines: Di-NeRF and RAMEN completely failed to reconstruct coherent scenes. MACIM produced incomplete maps with significant holes.
  • UDON: Successfully fused information to create complete, high-fidelity maps. It achieved the lowest Artifacts (4.40 cm vs. 7.15 cm for MACIM) and highest Completion Ratio (84.35% vs. 80.48% for MACIM) on the ScanNet dataset.
• Scalability (4 to 8 Agents):
  • UDON's performance improved as the number of agents increased, effectively leveraging complementary data even with sparse communication.
  • It maintained high fidelity in regions observed by only a few agents, whereas MACIM struggled in these sparse-coverage areas.
• Real-World Hardware (TurtleBot):
  • In a challenging 3-robot experiment with 5% communication success, UDON significantly outperformed MACIM in both artifact reduction and completion ratio.
  • The system handled the unique challenges of low-platform robots (obstructed views) by effectively fusing partial observations.
• Stability Analysis:
  • Loss curve analysis showed that Di-NeRF and RAMEN suffered from divergent dual variable losses, leading to training instability.
  • UDON stabilized the dual variable loss by isolating updates to active links, allowing for consistent convergence.

5. Significance

• Robustness in Critical Environments: UDON enables multi-robot teams to operate in environments where communication is severely degraded (e.g., subterranean exploration, disaster zones, or underwater missions), a scenario where previous neural mapping methods fail.
• Efficiency: By utilizing uncertainty weighting, the system maximizes the utility of every successful packet, ensuring that high-quality data drives the map reconstruction while low-quality data is down-weighted.
• Scalability: The framework scales effectively with the number of agents, proving that adding more robots improves map quality even when they cannot communicate frequently.
• Practical Application: The successful deployment on TurtleBot hardware validates that the method is not just a theoretical improvement but a viable solution for real-world robotic systems.

In conclusion, UDON represents a significant leap forward in collaborative robotics, solving the critical bottleneck of communication reliability in neural implicit mapping through a mathematically rigorous, uncertainty-aware distributed optimization approach.