Interaction-Merged Motion Planning

• IMMP is a checkpoint-merging framework that transfers diverse interaction patterns from multiple trajectory datasets, enhancing autonomous motion planning.
• It employs module-wise merging of ego, surrounding, and interaction encoders to preserve nuanced behaviors while mitigating domain imbalance and catastrophic forgetting.
• IMMP achieves competitive performance with single-model inference, outperforming ensembles in efficiency and adaptability across varied driving conditions.

Searching arXiv for the specified paper and closely related IMMP work. Interaction-Merged Motion Planning (IMMP) is a motion-planning-specific checkpoint-merging framework introduced for autonomous robot driving to transfer diverse interaction patterns from multiple trajectory datasets into a single target-domain planner. Rather than mixing raw datasets during adaptation or averaging whole models, IMMP constructs a pool of source-domain checkpoints, interprets them as task vectors from a common initialization, and learns how to merge them module by module—ego encoder, surrounding-agent encoder, interaction encoder, and remaining layers—using target-domain training loss. In that formulation, the central transferable object is not merely kinematic prediction accuracy but the structure of agent behavior and interaction encoded in planning models (Lee et al., 7 Jul 2025).

1. Problem setting and conceptual scope

IMMP is defined for data-driven motion planning models that take as input the ego agent’s past trajectory, surrounding agents’ past trajectories, and predicted future trajectories of surrounding agents, and output the ego agent’s future planned trajectory. The paper writes this as

$$X = \left\{\mathcal{X}_{ego}^{-T_{obs}:0},\ \mathcal{X}_{surr}^{-T_{obs}:0},\ \hat{\mathcal{Y}}_{surr}^{1:T_{fut}}\right\}, \qquad Y = \left\{\mathcal{Y}_{ego}^{1:T_{fut}}\right\}.$$

The objective is to use multiple source domains $\mathcal{S} = \{D_s^1,\dots,D_s^n\}$ to improve planning on target domains $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$ (Lee et al., 7 Jul 2025).

The motivation is that motion datasets differ not only in environmental geometry but in interaction structure. The cited source domains include ETH-UCY as human-human interaction data, THOR and SIT as human-robot interaction datasets, and CrowdNav as simulated robot navigation with ORCA humans and an RL robot. The paper identifies domain imbalance, catastrophic forgetting, interaction differences between datasets, environmental differences, computational cost, and the difficulty of knowing source usefulness in advance as the main obstacles to direct multi-domain training (Lee et al., 7 Jul 2025).

This makes IMMP distinct from standard domain adaptation, domain generalization, multitask learning, or ensembles. Domain-adaptation methods continue to expose the model to source data during target adaptation; ensembles retain multiple source models at inference; generic model-merging methods ignore the hierarchical interaction structure of planning models. IMMP instead treats source-domain checkpoints as reusable carriers of interaction knowledge and constructs a single merged model with inference cost $\times 1$, rather than the $\times 7$ cost reported for ensemble baselines (Lee et al., 7 Jul 2025).

A broader implication follows from the surrounding literature: interaction-aware planning is not a single algorithmic family but a design principle spanning intention-aware manipulation, traffic negotiation, shared-workspace planning, and contact-aware navigation. IMMP, in its named 2025 form, specializes that principle to checkpoint merging for trajectory planners, while earlier work more often merged interaction into prediction, cost design, or trajectory optimization directly. This suggests that the term denotes both a specific method and a wider interpretive frame, though the formal method is the checkpoint-merging framework of 2025 (Lee et al., 7 Jul 2025).

2. Two-stage method: pre-merging and interaction transfer

IMMP is organized as a two-stage pipeline. The first stage, “interaction-conserving pre-merging,” trains planners on each source domain and constructs a checkpoint pool $\mathcal{P}$. The second stage, “interaction transfer with merging,” learns how to combine those checkpoints on the target domain (Lee et al., 7 Jul 2025).

The checkpoint representation is task-vector based: $\tau_i = \Theta_i - \Theta_0,$ where $\Theta_0$ is the common initialization and $\Theta_i$ is a checkpoint in the pool. A global merged model is then written as

$$\Theta = \Theta_0 + \lambda \sum_{i=1}^{|\mathcal{P}|} w_i \cdot \tau_i.$$

The target-domain objective is

$\mathcal{S} = \{D_s^1,\dots,D_s^n\}$0

The paper states that $\mathcal{S} = \{D_s^1,\dots,D_s^n\}$1 and $\mathcal{S} = \{D_s^1,\dots,D_s^n\}$2 are fixed, and merging methods optimize the scaling and weight variables using target-domain train loss (Lee et al., 7 Jul 2025).

The distinctive element is that IMMP does not merge the model as a flat parameter vector. Instead it partitions parameters into interaction-relevant groups,

$\mathcal{S} = \{D_s^1,\dots,D_s^n\}$3

corresponding to the ego trajectory encoder, surrounding-agent encoder, interaction encoder, and decoder or remaining layers. The planning representation is described as

$\mathcal{S} = \{D_s^1,\dots,D_s^n\}$4

$\mathcal{S} = \{D_s^1,\dots,D_s^n\}$5

$\mathcal{S} = \{D_s^1,\dots,D_s^n\}$6

Each module is merged separately: $\mathcal{S} = \{D_s^1,\dots,D_s^n\}$7 That module-wise rule is the core technical distinction from ordinary task arithmetic (Lee et al., 7 Jul 2025).

Checkpoint selection is likewise planning-specific. For each source domain, IMMP stores metric-wise best checkpoints for ADE, FDE, collision rate, and miss rate, and also stores intermediate checkpoints every $\mathcal{S} = \{D_s^1,\dots,D_s^n\}$8 iterations. The stated rationale is that different metrics preserve different interaction characteristics, and intermediate checkpoints are often less overfit and therefore more transferable. The source-model update used to generate the pool is

$\mathcal{S} = \{D_s^1,\dots,D_s^n\}$9

where $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$0 denotes the backbone-specific planning loss (Lee et al., 7 Jul 2025).

3. Backbones, datasets, and empirical profile

IMMP is evaluated with three planning backbones: GameTheoretic, DIPP, and DTPP. It is therefore presented as backbone-agnostic rather than as a replacement architecture. The datasets span ETH-UCY, CrowdNav, THOR, and SIT. All are standardized to ego-centric $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$1 coordinates, sampled at $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$2 FPS, with $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$3 observed timesteps, $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$4 predicted future timesteps for surrounding agents, and a $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$5-step ego planning horizon. The principal target domains are SIT and THOR, chosen because they are real-world robotic datasets and relatively small (Lee et al., 7 Jul 2025).

The strongest reported result is on SIT with the GameTheoretic backbone. There, IMMP achieves ADE $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$6, CR $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$7, and FDE $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$8 at inference cost $\mathcal{T} = \{D_t^1,\dots,D_t^m\}$9, outperforming Ensemble-WTA at ADE $\times 1$0, CR $\times 1$1, and FDE $\times 1$2 with inference cost $\times 1$3. IMMP + Finetune further improves to ADE $\times 1$4, CR $\times 1$5, FDE $\times 1$6, and MR $\times 1$7 (Lee et al., 7 Jul 2025).

On DTPP with SIT as target, IMMP improves ADE to $\times 1$8, FDE to $\times 1$9, and MR to $\times 7$0, and IMMP + Finetune reaches ADE $\times 7$1 and FDE $\times 7$2. On DIPP with THOR as target, IMMP + Finetune reaches ADE $\times 7$3, CR $\times 7$4, FDE $\times 7$5, and MR $\times 7$6, improving substantially over Target Only at ADE $\times 7$7, FDE $\times 7$8, and MR $\times 7$9 (Lee et al., 7 Jul 2025).

The results are not uniformly monotone. On THOR, direct IMMP can be merely competitive or even weaker than Target Only or Domain Adaptation, particularly with GameTheoretic and DTPP. The consistent empirical pattern is that IMMP is a strong single-model method and IMMP + Finetune is often the best practical configuration. That is why the paper repeatedly characterizes IMMP as a strong initialization for target adaptation rather than as an unconditional replacement for target-only training (Lee et al., 7 Jul 2025).

The ablations sharpen the method’s internal logic. On GameTheoretic with SIT, interaction-level merging outperforms model-level and parameter-level merging, achieving ADE $\mathcal{P}$0 versus $\mathcal{P}$1 and $\mathcal{P}$2, respectively. Checkpoint-pool ablations show progressive improvement from ADE-only best checkpoints ($\mathcal{P}$3) to all-metric best checkpoints ($\mathcal{P}$4), then to all-metric plus intermediate planner checkpoints ($\mathcal{P}$5), and finally to the full forecaster/planner pool ($\mathcal{P}$6). The paper also reports that learned merging weights correlate qualitatively with source-target similarity: domains with better zero-shot target performance receive larger weights (Lee et al., 7 Jul 2025).

4. Relation to earlier interaction-aware planning

Although the named IMMP method is a checkpoint-merging framework, the broader interaction-aware planning literature had already developed several ways of merging interaction into motion planning. “I-Planner” predicts human actions and future motion with a progress-state-conditioned classifier-plus-GP model, then integrates probabilistic human occupancy directly into ITOMP replanning and task-level Q-learning; it is best described as prediction-augmented online motion planning rather than joint co-optimization (Park et al., 2016). “Interactive Model Predictive Control” for dense traffic learns a reference-velocity policy with SAC and feeds it into MPCC, so interaction enters planning through a low-dimensional guidance variable rather than through parameter merging (Brito et al., 2021).

Other works move closer to explicit multi-agent coupling. Branch MPC for AV-HV interaction optimizes over a scenario tree of human decision branches and enforces a shared plan prefix across branches until the human mode becomes observable; this is a clear instance of shared-prefix interaction-aware planning (Oliveira et al., 2023). “Interactive Joint Planning for Autonomous Vehicles” jointly optimizes ego and selected surrounding-agent trajectories while regularizing non-ego motions toward learned scene-centric prediction priors, making the trajectories of nearby agents decision variables in the optimizer (Chen et al., 2023). “Automated Vehicle Highway Merging” via adaptive interactive mixed-integer MPC jointly optimizes ego and neighboring-vehicle costs and updates the neighboring vehicle’s cost weights online through inverse optimal control (Bhattacharyya et al., 2022).

A separate line of work merges interaction through uncertainty-aware or constraint-coupled prediction. “Interaction-Aware Motion Planning for Autonomous Vehicles with Multi-Modal Obstacle Uncertainty Predictions” feeds maneuver probabilities and trajectory uncertainty into reference generation and MPC safety constraints through occupancy inflation (Zhou et al., 2022). “Interaction-Aware Vehicle Motion Planning with Collision Avoidance Constraints in Highway Traffic” couples predicted surrounding trajectories directly into an ego optimal-control problem with Pontryagin jump conditions, though only the ego is optimized (Kim et al., 2024). “A Risk and Comfort Optimizing Motion Planning Scheme for Merging Scenarios” jointly samples merge decisions and continuous trajectories instead of separating behavior and trajectory planning, thereby integrating tactical gap choice and motion generation under uncertainty (Müller et al., 2019).

Compared with those earlier methods, IMMP’s contribution is not a richer interaction dynamics model inside online planning. Its novelty lies in transferring interaction structure across datasets by checkpoint pooling and module-wise merging. In that sense, IMMP belongs to the interaction-aware planning lineage, but its direct object of manipulation is the planner’s parameter space rather than the online trajectory optimization itself (Lee et al., 7 Jul 2025).

5. Strengths, limitations, and methodological interpretation

The principal strength of IMMP is efficiency at deployment. It seeks to preserve the benefit of diverse source domains while keeping single-model inference cost $\mathcal{P}$7. This contrasts sharply with ensemble baselines, whose inference cost scales linearly with the number of component models. A second strength is that the method is explicitly planning-specific: checkpoint selection is metric-aware, merging is interaction-level rather than flat, and the grouping $\mathcal{P}$8 is chosen to respect the hierarchical structure of planning models (Lee et al., 7 Jul 2025).

A third strength is that IMMP does not require raw source data during the merging stage. Once the checkpoint pool is built, target adaptation uses fixed source checkpoints rather than continued multi-domain joint training. The paper interprets this as a way to reduce direct source-data competition, thereby mitigating domain imbalance and catastrophic forgetting. The results on SIT and on the larger Zara2 target domain support that interpretation: for Zara2, IMMP + Finetune reaches ADE $\mathcal{P}$9, FDE $\tau_i = \Theta_i - \Theta_0,$0, and MR $\tau_i = \Theta_i - \Theta_0,$1, surpassing Target Only and Ensemble-WTA while maintaining inference cost $\tau_i = \Theta_i - \Theta_0,$2 (Lee et al., 7 Jul 2025).

The limitations are equally explicit. Pre-merging is not free: source-domain models must still be trained and checkpointed. Gains are not uniform across backbones and targets; THOR provides several mixed-result cases where direct IMMP is not stronger than simpler baselines. The method also assumes that a meaningful decomposition into ego, surrounding, interaction, and other parameters exists, which is natural for LSTM- and Transformer-based trajectory planners but not guaranteed for every architecture. Finally, IMMP does not introduce a universal planning loss. It depends on each backbone’s $\tau_i = \Theta_i - \Theta_0,$3, so deployment still requires backbone-specific training infrastructure (Lee et al., 7 Jul 2025).

Methodologically, IMMP can be read as a planner-construction technique rather than a planner in the classical online-control sense. It does not alter the runtime semantics of GameTheoretic, DIPP, or DTPP. Instead it changes initialization and parameter composition so that interaction knowledge from heterogeneous domains is encoded in a single target planner. This suggests a layered interpretation of “interaction merging”: earlier work merged interaction into online rollout, MPC constraints, or game structure, whereas IMMP merges interaction at the level of learned planner parameters (Lee et al., 7 Jul 2025).

6. Terminological ambiguity and adjacent uses

The acronym “IMMP” is not unique. “MMP++: Motion Manifold Primitives with Parametric Curve Models” introduces “IMMP++,” but there the acronym denotes “Isometric Motion Manifold Primitives++,” not Interaction-Merged Motion Planning. IMMP++ concerns geometry-preserving latent spaces for motion primitives, continuous-time modulation, and online replanning in learned motion manifolds, and is unrelated to checkpoint merging across motion datasets (Lee, 2023).

Several papers are closely related to IMMP in spirit without instantiating the 2025 method. “IMPACT” turns VLM-inferred contact tolerances into dense 3D cost maps for contact-rich motion planning; it is contact-aware semantic motion planning rather than interaction-dynamics planning (Ling et al., 13 Mar 2025). “Beyond Task and Motion Planning” introduces Task and Skill Planning with Composable Interaction Primitives, wrapping closed-loop skills with motion-plannable entry and exit interfaces; this is a hierarchical skill-composition approach rather than parameter merging (Hedegaard et al., 24 Apr 2025). “A Reliable Robot Motion Planner in Complex Real-world Environments via Action Imagination” builds local contact models and imagined spatial states in cluttered environments, again merging interaction into online planning rather than into planner parameters (Wang et al., 21 Sep 2025).

There are also cases of bibliographic ambiguity. One cited 2020 record, despite its title and abstract, is described in the supplied technical details as a perception pipeline for unsupervised salient object discovery rather than an interaction-aware planner. In that reading it is only indirectly relevant to IMMP, as an upstream perception module that might provide novel object hypotheses to a downstream planner (Frank et al., 2020). Such ambiguity matters because “interaction-aware,” “integrated,” and “merged” planning now name a heterogeneous field: some methods merge interaction into online optimization, some into hierarchical skill interfaces, some into semantic cost maps, and IMMP proper merges it into reusable planner checkpoints.

Taken together, the literature supports a narrow and a broad definition. Narrowly, Interaction-Merged Motion Planning is the 2025 checkpoint-merging method for transferring interaction structure across planning datasets (Lee et al., 7 Jul 2025). Broadly, it names a family resemblance among planners that refuse to treat other agents, contact events, or human responses as fixed disturbances and instead incorporate them into the structure of planning itself.