SPREAD: Geometry-Preserving Lifelong Imitation Learning

• The paper introduces SPREAD, a framework that preserves low-dimensional manifolds across sequential tasks to mitigate catastrophic forgetting in lifelong imitation learning.
• It employs low-rank subspace alignment using singular value decomposition along with confidence-guided policy distillation to optimize multimodal representations.
• Empirical evaluations on the LIBERO benchmark demonstrate improved forward transfer, reduced forgetting, and state-of-the-art performance across diverse task suites.

SPREAD is a lifelong imitation learning framework that addresses catastrophic forgetting by preserving the geometry of multimodal task representations across sequential learning. It replaces raw feature matching with low-rank subspace alignment based on singular value decomposition and combines this with confidence-guided policy distillation over high-confidence action samples. In this formulation, the central object is not the coordinatewise feature vector itself but the low-dimensional manifold or subspace that encodes task structure across vision, language, and proprioception. The method is designed for sequential acquisition of skills from expert demonstrations, where a single policy must learn new tasks while retaining prior competence, and it is reported to improve knowledge transfer, mitigate catastrophic forgetting, and achieve state-of-the-art performance on the LIBERO lifelong imitation benchmark (Roy et al., 9 Mar 2026).

1. Lifelong imitation learning setting

SPREAD is formulated in the standard lifelong imitation learning setting in which tasks arrive sequentially as

$\{T^1, T^2, \dots, T^K\}.$

Each task $T^k$ is defined within a shared MDP $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$ by an initial state distribution $\mu_0^k$, a goal predicate $g^k : \mathcal{S} \to \{0,1\}$, and a language description $l^k$. For task $T^k$, the demonstration dataset is

$$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$

If all tasks were jointly available, behavior cloning would minimize

$$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$

The lifelong setting departs from this joint-training regime because data from past tasks are no longer fully available when learning new tasks; at best, a small replay buffer is retained. This induces the stability–plasticity dilemma: the policy must remain plastic enough to absorb new demonstrations while stable enough to preserve earlier skills. SPREAD targets this regime directly by assuming a teacher–student transition between successive policies, with $\pi^{k-1}$ frozen as teacher and $T^k$0 optimized on the current task together with replay (Roy et al., 9 Mar 2026).

The motivating claim is that multimodal policy embeddings lie on low-dimensional manifolds or subspaces that encode the essential structure of tasks. Standard feature distillation based on

$T^k$1

is described as sensitive to noise and high-dimensional variability, and as over-constraining because it enforces coordinatewise agreement rather than preserving intrinsic geometry. SPREAD therefore treats manifold geometry, rather than raw feature values, as the object that must persist under continual adaptation.

2. Subspace representation distillation

The representation module is built around modality-specific feature matrices extracted from the teacher and student policies. For a mini-batch of size $T^k$2 and sequence length $T^k$3, latent tensors are reshaped into matrices of size $T^k$4. The framework considers visual features $T^k$5 for modalities $T^k$6, text features $T^k$7, and extra features $T^k$8 for $T^k$9. The corresponding encoders are ResNet18 backbones for images, a CLIP language encoder followed by an MLP for text, and MLPs over joint and gripper states.

For a feature matrix $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$0, SPREAD applies a reduced SVD

$$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$1

and retains the top-$$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$2 left singular vectors. The associated projection operator is

$$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$3

The rank is fixed in the reported experiments at $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$4, described as roughly $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$5 of the full feature dimension. The ablation summary states that full rank retains noise and hurts forgetting, whereas overly small rank discards useful information; $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$6 gives the best trade-off between transfer and forgetting on LIBERO-GOAL (Roy et al., 9 Mar 2026).

The core subspace alignment loss between teacher features $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$7 and student features $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$8 is

$$\mathcal{M} = (\mathcal{S}, \mathcal{A}, \mathcal{T}, H)$$9

This objective is symmetric. The first term aligns the projected content of teacher and student in their respective low-rank subspaces, while the second cross term encourages the teacher and student subspaces themselves to align. The paper characterizes this as geometry-aware distillation because it emphasizes principal directions rather than coordinatewise equality.

SPREAD applies this construction independently to each modality: $\mu_0^k$0

$\mu_0^k$1

$\mu_0^k$2

This modality-wise decomposition is important because the policy is multimodal: semantic language, egocentric and exocentric vision, and proprioception are all treated as distinct representational channels whose low-rank geometry should remain stable under sequential training.

A plausible implication is that the method constrains only the task-relevant directions of variation while leaving orthogonal directions available for new-task adaptation. The paper motivates this directly by noting that low-rank projections filter components aligned with small singular values and therefore suppress noise and high-dimensional variability.

3. Confidence-guided policy distillation and total objective

Representation preservation is coupled with a second mechanism that distills action distributions. The policy outputs a Gaussian mixture model over actions, and SPREAD does not apply a full Monte Carlo KL over all sampled actions. Instead, it samples $\mu_0^k$3 actions

$\mu_0^k$4

computes their log-probabilities under the old policy, selects the top

$\mu_0^k$5

most confident samples, and restricts the distillation loss to that subset.

The resulting objective is

$\mu_0^k$6

This is described as equivalent up to constants to a Monte Carlo estimate of

$\mu_0^k$7

restricted to high-density regions under the teacher policy. The operational rationale is that low-likelihood samples are uncertain or uninformative modes; excluding them reduces variance and emphasizes reliable action structure.

During task $\mu_0^k$8, the student is trained on the current demonstrations $\mu_0^k$9 together with a replay buffer $g^k : \mathcal{S} \to \{0,1\}$0. The full objective is

$g^k : \mathcal{S} \to \{0,1\}$1

The behavior cloning term supplies plasticity on new data, while the subspace and policy losses supply stability. SPREAD therefore does not replace replay; it augments replay with geometry-preserving regularization and confidence-guided action matching (Roy et al., 9 Mar 2026).

The training procedure is sequential and teacher–student. For each new task, $g^k : \mathcal{S} \to \{0,1\}$2 is frozen, $g^k : \mathcal{S} \to \{0,1\}$3 is initialized from it, feature matrices are extracted for both policies, SVD is performed per modality, subspace and policy losses are computed, and the total loss is backpropagated. After optimization, exemplars from the new task are added to replay.

4. Empirical performance on LIBERO

SPREAD is evaluated on LIBERO, described as a language-conditioned manipulation benchmark specifically designed for lifelong imitation learning. The reported experiments use three 10-task suites: LIBERO-OBJECT, LIBERO-GOAL, and LIBERO-SPATIAL. These suites test, respectively, object-focused transfer, goal-oriented procedural knowledge, and spatial relational understanding. The observation space includes HandEye and AgentView vision, language instructions, and proprioception.

The principal metrics are Forward Transfer (FWT), Negative Backward Transfer (NBT), and AUC, all reported as percentages. FWT measures success when each task is first learned, NBT measures performance drop on earlier tasks after subsequent learning and is lower-is-better, and AUC measures overall success over the full learning trajectory.

Before the table, two patterns are worth emphasizing. First, SPREAD is reported to improve overall performance across all three suites, especially on the harder LIBERO-GOAL benchmark. Second, the gains are not only in forward transfer; they are also in forgetting, which is the central failure mode of lifelong imitation learning.

Suite	SPREAD	Representative baseline
LIBERO-OBJECT	FWT $g^k : \mathcal{S} \to \{0,1\}$4, NBT $g^k : \mathcal{S} \to \{0,1\}$5, AUC $g^k : \mathcal{S} \to \{0,1\}$6	M2Distill: FWT $g^k : \mathcal{S} \to \{0,1\}$7, NBT $g^k : \mathcal{S} \to \{0,1\}$8, AUC $g^k : \mathcal{S} \to \{0,1\}$9
LIBERO-GOAL	FWT $l^k$0, NBT $l^k$1, AUC $l^k$2	M2Distill: FWT $l^k$3, NBT $l^k$4, AUC $l^k$5
LIBERO-SPATIAL	FWT $l^k$6, NBT $l^k$7, AUC $l^k$8	M2Distill: FWT $l^k$9, NBT $T^k$0, AUC $T^k$1

On LIBERO-OBJECT, SPREAD improves FWT by $T^k$2 and AUC by $T^k$3 relative to M2Distill, with the same NBT. On LIBERO-GOAL, the improvement is substantially larger: $T^k$4 FWT relative to M2Distill, $T^k$5 NBT, and $T^k$6 AUC. On LIBERO-SPATIAL, SPREAD has slightly lower FWT than M2Distill but lower forgetting and higher AUC. The aggregate interpretation in the paper is that SPREAD consistently yields better overall performance and less forgetting (Roy et al., 9 Mar 2026).

The trajectory plots described in the paper support the same conclusion. Success-rate curves remain high and stable across sequential tasks, forward transfer remains comparatively stable on later tasks, and NBT stays low on LIBERO-GOAL while M2Distill shows substantially larger positive NBT.

5. Ablations, rank selection, and representation stability

The ablation studies isolate the contributions of the distinct losses. On LIBERO-GOAL, removing $T^k$7 reduces AUC from $T^k$8 to $T^k$9, indicating that text alignment helps but is not dominant. Removing $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$0 is much more damaging, reducing AUC from $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$1 to $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$2 and increasing NBT from $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$3 to $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$4. This identifies visual subspace preservation as the principal contributor in the reported setting. Removing $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$5 lowers FWT from $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$6 to $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$7, which the paper interprets as evidence that action distillation materially supports transfer even though forgetting can decrease slightly. Removing $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$8 mildly hurts both FWT and AUC.

Rank selection is also explicitly studied. On LIBERO-GOAL, full-rank distillation yields FWT $$D^k = \{\tau_i^k\}_{i=1}^N,\qquad \tau_i^k = \{(s_t,a_t)\}_{t=1}^{L^k},\qquad L^k < H.$$9, NBT $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$0, and AUC $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$1; rank $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$2 yields FWT $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$3, NBT $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$4, and AUC $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$5; rank $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$6 yields FWT $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$7, NBT $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$8, and AUC $$\min_{\pi} J(\pi) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{E}_{(s_t,a_t)\sim D^k} \left[\sum_{t=1}^{L^k} -\log \pi(a_t\mid s_{\le t}; T^k)\right].$$9; rank $\pi^{k-1}$0 yields FWT $\pi^{k-1}$1, NBT $\pi^{k-1}$2, and AUC $\pi^{k-1}$3. The best overall performance is therefore obtained by a moderate low-rank subspace, rather than by full-rank matching or aggressive compression. This supports the paper’s argument that an appropriate low-rank approximation captures essential manifold structure while filtering nuisance variation.

The confidence-guided policy ablation leads to the same kind of conclusion. Using the top-$\pi^{k-1}$4 samples is reported as the best setting, outperforming both top-$\pi^{k-1}$5 and all-sample variants in AUC and NBT. This suggests that distillation should be concentrated on high-confidence modes of the teacher policy rather than spread across low-density action regions.

A separate representation-drift analysis quantifies the geometry-preserving claim. On LIBERO-OBJECT, language embedding drift is reduced by more than $\pi^{k-1}$6 relative to M2Distill. For HandEye vision, SPREAD keeps drift below $\pi^{k-1}$7, whereas M2Distill reaches above $\pi^{k-1}$8. For AgentView vision, mean drift is $\pi^{k-1}$9 for SPREAD versus $T^k$00 for M2Distill. This does not by itself prove manifold preservation, but it is direct evidence that sequential representation changes are substantially smaller under the SPREAD objective (Roy et al., 9 Mar 2026).

6. Limitations and position within continual learning research

SPREAD is presented as a geometry-preserving alternative to raw-feature distillation, but its design also imposes several explicit limitations. First, it incurs the computational cost of SVD for each modality and batch, with complexity $T^k$01 per modality. The paper describes this as acceptable at current scales but potentially significant for larger models or many modalities. Second, the subspace rank $T^k$02 is fixed and tuned empirically; adaptive rank selection is left open. Third, experiments are confined to 10-task suites, so scaling to much longer task sequences may require additional mechanisms for replay management or architecture adaptation. Fourth, the framework remains purely imitation-based and therefore depends on high-quality expert demonstrations. Finally, the method approximates multimodal manifolds with linear subspaces, which may be restrictive when the true geometry is strongly nonlinear.

Within the broader continual learning landscape, SPREAD is positioned against parameter regularization, replay, and raw-feature distillation. The relevant contrast is not between replay and non-replay methods, since SPREAD still uses a replay buffer, but between coordinatewise feature preservation and geometry-preserving subspace preservation. The paper situates the method alongside knowledge distillation and geometry-aware continual learning, and describes its novelty as modality-wise subspace distillation in lifelong imitation learning, a symmetric subspace loss that aligns both self and cross projections, and confidence-guided policy distillation for Gaussian mixture policies.

A common misconception would be to interpret SPREAD as a generic replacement for all continual learning mechanisms. The framework, as specified, is more precise: it is a teacher–student distillation method for multimodal lifelong imitation learning that preserves low-rank geometry and high-confidence action modes while relying on replay and behavior cloning for the rest of the optimization. Another misconception would be to read the subspaces as exact task manifolds. The paper does not claim exact manifold recovery; it uses linear low-rank subspaces as tractable approximations to the low-dimensional structure of multimodal embeddings.

SPREAD therefore occupies a specific methodological niche. It treats catastrophic forgetting as, in part, a geometric instability problem in representation space, and it addresses that problem with SVD-based subspace alignment plus confidence-restricted action matching. This suggests a broader research direction in which continual robot learning is analyzed not only through parameter drift or replay coverage, but through the preservation of latent geometry across sequential tasks (Roy et al., 9 Mar 2026).