Experience-Imitation Planning Insights

• Experience-Imitation Planning is a strategy that reuses previous planning episodes by converting trajectories, demonstrations, and symbolic traces into guidance for future control.
• It employs a retrieve–adapt–compose cycle along with planning-conditioned supervision and search-control learning to optimize decision-making in dynamic environments.
• Applications span multi-robot coordination, motion planning, and autonomous driving, achieving notable improvements in success rates, memory usage, and computational speed.

to=arxiv_search.search 天天爱彩票网站 code meant maybe tool name? In the literature surveyed here, Experience-Imitation Planning can be understood as an umbrella term for methods that convert prior solved trajectories, demonstrations, observation sequences, symbolic task structure, or previous planning episodes into guidance for future planning and control. Across these works, the reusable object is not fixed: it may be a full path in a sparse roadmap, a conflict-localized multi-robot coordination fragment, a learned reward over future observations, a symbolic operator trace inferred by abduction, an oracle planner’s action choice, a self-generated demonstration synthesized by planning over visited states, or a dense latent future trajectory used for receding-horizon control (Coleman et al., 2014, Solis et al., 2024, Jaegle et al., 2021, Shao et al., 2024, Boney et al., 2020, Luo et al., 2021, Xie et al., 23 Apr 2025). A plausible implication is that the field is unified less by a single algorithmic template than by a recurring design principle: prior successful computation is retained and reused inside later planning loops rather than discarded after training or execution.

1. Conceptual scope and reusable planning objects

A central issue in Experience-Imitation Planning is the identity of the reusable unit. In continuous multi-robot motion planning, the reusable object is a solved local multi-robot subproblem rather than a full-team solution: E-ARC stores local coordination patterns for “how 2–4 robots coordinate through a local interaction,” then retrieves, transforms, validates, optionally repairs, and inserts them into a larger plan (Solis et al., 2024). In Thunder, by contrast, the reusable object is not an individual path but a graph stored in a Sparse Roadmap Spanner, so that subpaths can be recombined under changing obstacles while preserving asymptotically near-optimal coverage of configuration space (Coleman et al., 2014). In ERT and IERTC*, the reusable unit is even more local: a prior path is treated as “decomposable” and “malleable,” split into micro-paths and morphed online to generate new local motions (Pairet et al., 2021, Takamido et al., 19 Mar 2025).

In demonstration-based methods, the reusable object shifts from geometry to supervision. FORM replaces expert actions with a predictive density over observation transitions, using

$$r_t=\log p_\omega^D(x_t\mid x_{t-1})-\log p_\phi^I(x_t\mid x_{t-1}),$$

so that expert experience becomes a reward signal for control without action labels (Jaegle et al., 2021). ABIL reuses demonstrations in a different way: demonstrations are abductively interpreted as latent symbolic traces consistent with a finite-state-machine knowledge base, and those inferred traces supervise both perception and operator-specific behavior modules (Shao et al., 2024). In planner-distillation settings, the reusable object is an oracle planner’s decision: a full-state “cheating AI” generates actions that a partially observed follower policy later imitates under DAgger-style aggregation (Boney et al., 2020).

This variation in reusable object is not incidental. It suggests that experience-imitation methods succeed when they preserve the level of abstraction at which structure repeats. Some domains repeat local geometric interactions; some repeat symbolic operator orderings; some repeat observation-transition regularities; some repeat planner decisions under privileged information.

2. Recurrent architectural patterns

Across the surveyed papers, several architectural motifs recur. The first is retrieve–adapt–compose. E-ARC retrieves the $k$ closest stored subproblem solutions, aligns them to the current conflict, checks endpoint connectability and collisions, optionally repairs them, and composes them into the global path set (Solis et al., 2024). Thunder retrieves graph paths from a SPARS experience graph, lazily invalidates inconsistent edges, and repairs the best partial route with bidirectional RRT when necessary (Coleman et al., 2014). ERT and IERTC* similarly retrieve one prior trajectory, morph local fragments, then compose them through tree growth and rewiring (Pairet et al., 2021, Takamido et al., 19 Mar 2025).

A second motif is plan with an imitation-derived evaluator. FORM does not perform explicit online search, but it still turns demonstrations into a trajectory-scoring mechanism over future observations and then uses RL to optimize long-horizon behavior under that score (Jaegle et al., 2021). IMPLANT keeps both the imitation policy and the imitation-derived reward model at test time, performs MPC-style rollouts, and selects actions by maximizing

$$\hat{R}_{\phi,\psi}(\tau)=\sum_{t'=t}^{t+H-1}\gamma^{t'-t}r_\phi(s_{t'},a_{t'})+\gamma^H V_\psi(s_H),$$

thereby using demonstrations twice: once to learn proposals and once to learn online evaluation (Qi et al., 2022). MPAIL goes further by replacing the reactive policy in adversarial imitation learning with MPPI itself, so that the planner is inside the imitation loop rather than downstream of it (Han et al., 29 Jul 2025).

A third motif is planning-conditioned supervision generation. SILP and SILP+ collect visited states from the current policy, reinterpret them as planning nodes, synthesize demonstrations with PRM-style planning, and then train the policy with a behavior-cloning auxiliary loss gated by an advantage-like filter (Luo et al., 2021, Luo et al., 2023). In these systems, planning is not the deployed controller; it is a demonstration generator that continuously converts the agent’s own experience into new imitation data.

A fourth motif is search-control learning rather than trajectory learning. “Leveraging Experience in Lazy Search” formulates edge evaluation ordering in LazySP as an MDP over search states and learns, by imitation of an oracle, which edge to evaluate next (Bhardwaj et al., 2021). “Experience-driven discovery of planning strategies” studies not external action choice but the acquisition of internal planning procedures—what to inspect, in what order, and when to stop—through metacognitive reinforcement learning (He et al., 2024). A plausible implication is that Experience-Imitation Planning includes methods that learn how to allocate computation, not only how to allocate motion.

3. Representative paradigms

The surveyed literature organizes naturally by the form of retained experience and the role that experience plays at planning time.

Form of retained experience	Planning role	Representative papers
Full paths, graphs, subproblems, or micro-paths	Retrieval, transformation, repair, and composition inside geometric planners	(Coleman et al., 2014, Solis et al., 2024, Pairet et al., 2021, Takamido et al., 19 Mar 2025)
Observation-transition densities or adversarial rewards	Trajectory scoring for RL, MPC, or MPPI	(Jaegle et al., 2021, Qi et al., 2022, Han et al., 29 Jul 2025)
Symbolic traces and operator-conditioned demonstrations	Long-horizon symbolic planning with learned perception and subpolicies	(Shao et al., 2024)
Oracle planner outputs under privileged information	DAgger-style planner distillation into deployable followers	(Boney et al., 2020)
Planner-generated self-demonstrations from visited states	Auxiliary imitation supervision for off-policy RL	(Luo et al., 2021, Luo et al., 2023)
Dense latent future trajectories	Receding-horizon latent planning with inverse dynamics execution	(Xie et al., 23 Apr 2025)
Edge-selection or metacognitive search traces	Learned control of search effort and internal planning strategy	(Bhardwaj et al., 2021, He et al., 2024)

This taxonomy suggests that “experience” can enter a planner as memory, reward, abstraction, proposal distribution, self-generated supervision, or control over computation. The papers differ most strongly in where that experience is inserted: before planning as a prior, during planning as a retrieval/evaluation mechanism, or after planning as a distilled policy.

4. Geometric motion planning: from path memories to subproblem imitation

Experience-based motion planning provides one of the clearest genealogies for the topic. Thunder stores experiences in a SPARS2 graph rather than as individual paths, allowing subpath recombination, lazy collision checking, and bounded-growth behavior. On a 30-DOF humanoid, it was reported to be “on average an order of magnitude faster than Lightning and planning from scratch” and to use 98.8% less memory after 10,000 trials (Coleman et al., 2014). This line of work shows that a path library becomes more reusable when its structure is factorized.

E-ARC extends that intuition from single-robot experience reuse to continuous-space MRMP. Its core claim is that naive full-problem databases do not scale because database construction, retrieval, and repair all become expensive as team size grows. The method instead stores low-dimensional conflict-localized coordination patches. In experiments it used a low-dimensional database built once in 7.42 s for mobile robots and 19.12 s for manipulators, and reported experimental validation with up to 32 mobile and 16 manipulator robots (Solis et al., 2024). The resulting planner is best viewed as local experience-guided repair inside ARC’s adaptive conflict-resolution loop.

ERT, ERTConnect, and IERTC* push the decomposition further. ERT introduces the claim that experiences are “decomposable” and “malleable”: instead of repairing whole retrieved paths, the planner incrementally composes a new motion from morphed micro-experiences extracted from a single prior path (Pairet et al., 2021). IERTC* then adds rewiring and informed pruning. Its reported results emphasize difficult clutter: with one hundred experiences it improved success rate by an average of 49.3% and reduced solution cost by 56.3%, while even with one experience it still reported 43.8% improvement in success rate and 57.8% reduction in solution cost (Takamido et al., 19 Mar 2025).

A common misconception is that all experience-based motion planners merely replay or patch whole trajectories. These papers show otherwise. The reusable object can be a sparse graph, a conflict-localized coordination patch, or a morphed path fragment. Composition is typically heuristic and search-based rather than guaranteed by an exact decomposition theorem, but it is precisely that local compositionality that permits reuse under changed environments.

5. Demonstrations as rewards, planners, and self-generated supervision

A second major lineage uses demonstrations not primarily as paths to retrieve, but as signals from which planners or evaluators are built. FORM is exemplary: it addresses imitation from observation by learning demonstrator and imitator effect models and then defining reward through their log-likelihood difference (Jaegle et al., 2021). It is therefore “a reward-learning-based, observation-only imitation method that can support planning/control,” not an explicit online planner. This distinction matters: experience-imitation planning includes indirect planning mechanisms in which demonstrations determine a control objective rather than a trajectory library.

IMPLANT and MPAIL make planning explicit. IMPLANT keeps the imitation policy and learned reward model at test time and performs decision-time planning to correct compounding errors of the base imitator (Qi et al., 2022). MPAIL goes further by replacing the policy generator in adversarial imitation learning with an MPC agent, using transition-level discriminator rewards and a learned value function for infinite-horizon MPPI (Han et al., 29 Jul 2025). In real-world navigation from a single partially observable demonstration, MPAIL reported mean cross-track error 0.17 m versus 0.37 m for IRL-MPC and 0.56 m for GAIL (Han et al., 29 Jul 2025). A plausible implication is that planner-in-the-loop imitation changes not only deployment but also the reward landscape seen during learning.

Planner distillation provides a third pattern. In imperfect-information games, a full-state oracle planner built with fixed-depth tree search and decoupled Thompson sampling produces high-quality privileged decisions that a partial-information follower then imitates with DAgger (Boney et al., 2020). The follower was trained on 500 battles in Pommerman and 300 battles in Clash Royale, illustrating a sample-efficiency story very different from standard model-free RL (Boney et al., 2020).

Self-imitation by planning provides a fourth pattern. SILP and SILP+ collect collision-free or feasible visited states, build planning graphs over those states, generate demonstrations automatically, and then train the actor with an auxiliary behavior-cloning loss (Luo et al., 2021, Luo et al., 2023). In SILP+, the reported best result was SAC-SILP+ with success rate $0.973 \pm 0.002$, compared to plain SAC at $0.864 \pm 0.066$, together with extensive physical-robot evaluation (Luo et al., 2023). These methods are especially notable because “experience” here is the learner’s own imperfect rollouts, transformed by planning into imitation targets.

6. Long-horizon abstraction and closed-loop deployment

Long-horizon planning introduces a different challenge: raw observation-action imitation degrades as temporal error compounds. ABIL addresses this by inserting an abductive symbolic layer between demonstrations and control. Demonstrations are interpreted as latent symbolic trajectories constrained by a finite-state machine, perception is trained from abduced pseudo-labels, and a policy ensemble of operator-specific modules is gated by symbolic reasoning (Shao et al., 2024). In Mini-BEHAVIOR, the tasks of interest span average lengths from roughly 10 up to 106 primitive decisions, and the strongest evidence for planning competence comes from zero-shot composition and generalization results such as Unlock and repeated leftover-disposal tasks (Shao et al., 2024). This suggests that experience-imitation planning can be hierarchical and neuro-symbolic rather than purely geometric or reward-based.

Autonomous driving provides two additional variants. PILOT trains a neural planner to imitate an expensive optimization-based expert, then refines the neural proposal online with an optimizer solving the same objective. On CARLA benchmarks it reported a seven-fold reduction in runtime without sacrificing planning quality (Pulver et al., 2020). The contribution is not mere policy distillation; it is learning a warm start for constrained optimization while retaining the model-based correction layer online.

A different systems-level lesson appears in the closed-loop framework for imitation-based joint prediction and planning. That work argues that high-frequency replanning is not always beneficial and that a learned prediction-planning model should operate in two modes: Motion Prediction and Planning (MPP) at a lower planning frequency, and Conditional Motion Prediction (CMP) as a higher-frequency safety monitor (Guo et al., 2024). On nuPlan Val14, the reported scores were 88.9 NR-CL and 84.0 R-CL with adaptive scheduling, compared with 84.8 and 76.8 for PlantTF (Guo et al., 2024). The significance is deployment-centric: experience-imitation planning includes not only how a planner is trained, but also how an imitation-trained planner is executed in feedback with the environment.

Latent Diffusion Planning adds a recent generative variant. It separates latent trajectory planning from inverse dynamics, allowing action-free trajectories to supervise the planner and suboptimal action-labeled data to supervise the inverse model (Xie et al., 23 Apr 2025). In low-demo visual manipulation, the combined LDP + Action-Free + Subopt average of 0.95 exceeded baselines that could not naturally exploit both data sources (Xie et al., 23 Apr 2025). Here the reusable experience is a dense future latent trajectory rather than an action trace.

7. Limitations, misconceptions, and open directions

Several limits recur across the surveyed works. First, many methods rely on local similarity assumptions. E-ARC assumes most conflicts involve only a few robots; if coordination groups become large, database reuse weakens (Solis et al., 2024). ERT-family planners assume useful micro-path structure transfers across tasks; if the required homotopy class changes drastically, prior fragments may become poor guides (Pairet et al., 2021, Takamido et al., 19 Mar 2025). Lazy-search imitation assumes future problems are sufficiently similar to training worlds that edge-selection policies transfer (Bhardwaj et al., 2021).

Second, guarantees are often inherited or conditional rather than intrinsic. E-ARC’s effective completeness comes from fallback planning rather than database retrieval alone (Solis et al., 2024). PILOT’s safety claims are conditional on successful convergence of the online optimizer (Pulver et al., 2020). The closed-loop prediction-planning framework argues for feasibility and local stability empirically rather than through formal theorems (Guo et al., 2024). MPAIL depends on model quality, and vanilla MPPI remains difficult in high-dimensional spaces (Han et al., 29 Jul 2025).

Third, the literature repeatedly shows that not all imitation-based planners are explicit planners. FORM supports planning/control indirectly through reward learning (Jaegle et al., 2021). ABIL is planner-like because symbolic reasoning selects operator-specific policies, but primitive-action search is not performed online (Shao et al., 2024). By contrast, IMPLANT, MPAIL, PILOT, EI, and LDP all retain explicit decision-time planning components (Qi et al., 2022, Han et al., 29 Jul 2025, Pulver et al., 2020, Yin et al., 2022, Xie et al., 23 Apr 2025). Treating these as identical would obscure an important methodological division.

Fourth, human strategy discovery suggests current computational accounts remain incomplete. Metacognitive RL models can discover planning strategies, but when fit to human data they still exhibit a slower discovery rate than humans (He et al., 2024). A plausible implication is that future Experience-Imitation Planning may need stronger mechanisms for abstraction, memory, or self-imitation of successful internal computations, rather than incremental reinforcement alone.

Taken together, these works indicate that Experience-Imitation Planning is best seen as a family of methods for retaining, abstracting, and reusing prior successful computation inside later planning loops. The strongest recurring lesson is that the reusable object must match the domain’s natural invariants: local coordination patches in MRMP, sparse graph structure in geometric planning, latent symbolic traces in long-horizon tasks, planner actions under privileged information in games, self-generated demonstrations in robot learning, and transition-level occupancy rewards in planner-in-the-loop imitation. Where that match is well chosen, experience becomes not just data for training, but an active computational resource at decision time.