Unified Prediction and Planning via Conflict-Aware Disjoint Parameter Training

Taewon Seo, Seonae Jeon, Giwon Lee, Kuk-Jin Yoon, Daehee ParkarXiv 2026 · arXiv

TL;DR

Compact unified models for robot navigation often suffer from "Skill Conflict," where motion prediction and safety planning compete for the same weights in a shared encoder. This paper introduces Disjoint Parameter Training (DPT) and Sparse Merging, a framework that isolates task-specific parameter regions during training and combines only the most critical weights. The result is a high-performance, resource-efficient model perfectly optimized for deployment on edge devices.

What Problem Does This Paper Address?

Social robots navigating crowded environments need to do two things simultaneously: predict how nearby humans will move (motion prediction) and plan a safe path for themselves (motion planning). To run efficiently on resource-constrained edge devices, state-of-the-art systems use a single, compact shared encoder to handle both tasks at once.

However, these two tasks have fundamentally opposing goals. Prediction requires high sensitivity to diverse, varied external behaviors, while planning demands ego-centric stability and safety features. Forcing a compact network to optimize both objectives simultaneously leads to conflicting gradient updates. The authors term this Skill Conflict—a phenomenon where overlapping parameter assignments cause distinct tasks to cannibalize the same weights, severely degrading the feature quality and performance of both components.

Methods

To eliminate Skill Conflict without inflating the model size, the authors propose a novel model-merging-based framework called Disjoint Parameter Training (DPT), followed by Sparse Merging.

Fig. 1: Comparison of learning strategies in unified models. (a) In previous methods, the skills of each task are learned within a single shared encoder, causing conflicts be- tween the two skills in the parameter space. (b) DPT resolves this issue by separately training the major parameter regions allocated to each task and Sparse Merging com- bines only their core parameter regions, thereby mitigating skill conflicts within the shared encoder.
  1. Disjoint Parameter Training (DPT): Instead of training both tasks jointly from the start, DPT prepares task-specialized "material models" by dynamically allocating non-overlapping parameter regions. It alternates training between the prediction and planning tasks. At each step, it identifies unassigned parameters with the largest absolute gradients for the current task, permanently locks them in using a binary mask, and updates only those weights. These masks grow incrementally during the first half of training and freeze during the second half to consolidate core skills.
  2. Sparse Merging: Even when parameters are separated into disjoint masks, adjacent weights can cause functional feature interference. To fix this, the authors introduce Sparse Merging, which combines the fine-tuned models using only a highly restrictive fraction (e.g., the top 1% by magnitude) of each task's most influential weights.
  3. Joint Reasoning Phase: To preserve the inherent benefits of cross-task interaction, a final fine-tuning phase is introduced using a joint task loss on the parameters that remained inactive during the DPT phase.

Datasets

The framework was evaluated on standard crowd navigation benchmarks:

  • JRDB: A real-world dataset featuring indoor and outdoor pedestrian trajectories captured by a mobile robot, complete with annotated bounding boxes.
  • JTA: A large-scale synthetic dataset providing rich annotations for pedestrian motion and 3D keypoints (the authors utilized a 15% subset to maintain a comparable experimental scale).
  • Bench2Drive: Used to demonstrate the scalability of the method in an end-to-end autonomous driving setup within the vehicle domain.

Key Insights and Ideas

  • The Capacity Paradox: In massive foundation models, tasks naturally find their own sparse, non-interfering subspaces. In compact edge-deployable models, however, tasks greedily expand into the same parameter regions, causing severe overlap (surging up to ~40% overlap at a 50% mask ratio), which triggers early performance saturation.
  • Spatial Disjointness Isn't Enough: Simply separating parameters into non-overlapping sets during training still allows functional interference if the updates occur in adjacent coordinates or nearby layers. Sparse Merging is essential to isolate footprints and prevent feature coupling.
  • Pervasiveness of Skill Conflict: Through extensive sweeps, the authors proved that Skill Conflict is not specific to one architecture; it consistently appears across multiple backbones (such as Social-Transmotion and DTPP) and intensifies as model dimensions shrink.

Key Results

  • Substantial Performance Gains: On the JRDB dataset, combining DPT with Sparse Merging and Joint Reasoning (DPT + SM + JR) achieved a planning Average Displacement Error (ADE) of 0.3640, drastically outperforming standard unified baselines like DIPP (0.8048) and DTPP (0.6232).
  • Edge-Ready Efficiency: The final model is incredibly lightweight, containing just 6.44M parameters (a mere 24.55 MB model size). It operates with a latency of 10.4 ms and requires only 1.65 GFLOPs, making it fully deployable on low-power edge hardware like the NVIDIA Jetson Orin Nano.
  • Broad Compatibility: DPT is broadly applicable; incorporating it into the training phase of other popular model-merging techniques (such as Task Arithmetic and Ties Merging) consistently yielded overall performance improvements.

Limitations and Future Work

  • Scaling Beyond Two Tasks: The current formulation focuses primarily on a 2-way split (prediction vs. planning). Extending DPT to broader multi-task settings with a larger number of diverse objectives will require more complex, principled task grouping and parameter allocation designs.
  • Hyperparameter Sensitivity: The framework relies on manual tuning for parameters like the allocation ratio, mask ratio, and task training order. Developing adaptive or automated selection methods for these hyperparameters remains an important direction for future research.

Summary updated

Abstract

Accurate motion prediction of surrounding agents and safe motion planning are two closely coupled key tasks for social robot navigation in crowded environments. Deploying these systems on resource-constrained edge devices necessitates compact, unified models that can perform both tasks simultaneously. However, within these compact shared encoders, recent unified models often overlook severe representational conflicts that arise from the distinct objectives of predicting neighbor behaviors versus ego-centric safety planning. To address this issue, we first identify the Skill Conflict$\unicode{x2014}$a phenomenon where overlapping parameter assignments cause distinct tasks to compete for the same weights, preventing the model from fully specializing in individual skills. To resolve this, we propose a novel model-merging-based framework, Disjoint Parameter Training (DPT). DPT mitigates performance degradation caused by Skill Conflict through distributed parameter learning, which separates the key parameter regions of each task while preserving their core capabilities prior to merging. In addition, we observe that sparse merging, which selectively integrates only the most influential parameters for each task rather than combining all task-specific parameters, yields optimal performance by preventing interference among adjacent features and concentrating representational capacity. DPT can be applied in parallel with a variety of merging methods. Evaluated on standard crowd navigation benchmarks (JRDB and JTA), our framework demonstrates superior performance, validating its versatility and effectiveness for safe, resource-efficient robot navigation.
BibTeX
@misc{seo2026unified,
  author = {Taewon Seo and Seonae Jeon and Giwon Lee and Kuk-Jin Yoon and Daehee Park},
  title = {{Unified Prediction and Planning via Conflict-Aware Disjoint Parameter Training}},
  year = {2026},
  eprint = {2607.19971},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.19971},
}
Figure