LIBERO-CF: Coarse-to-Fine Action Generation

• LIBERO-CF is an ambiguous term representing a coarse-to-fine action generation mechanism within the LIBERO benchmark, highlighting structured initialization for efficient inference.
• It is evaluated in CF-VLA by decomposing action generation into a coarse initialization stage and a single-step local refinement that drastically reduces function evaluations and action latency.
• The term’s usage varies, with interpretations as collision-free subsets and counterfactual failures, underscoring its non-canonical status in contemporary Vision-Language-Action literature.

LIBERO-CF is an ambiguous term in the recent Vision-Language-Action literature rather than a uniformly standardized benchmark name. In the clearest documented usage among currently cited arXiv sources, it denotes the coarse-to-fine setting introduced by CF-VLA on the LIBERO manipulation benchmark, where action generation is restructured into a coarse initialization step followed by a single-step local refinement (Du et al., 27 Apr 2026). At the same time, later LIBERO-family papers use the label differently or question its benchmark status: one source states that “LIBERO-CF (Collision-Free) is not an established official LIBERO benchmark,” while another describes “LIBERO-CF” as “Counterfactual Failures, Fang et al. 2026” in contrast to language-only perturbation studies (Cui et al., 22 Jun 2026, Kim et al., 30 Mar 2026). This suggests that, in current usage, LIBERO-CF is best understood through its surrounding context: most concretely as the LIBERO evaluation setting for CF-VLA’s coarse-to-fine action generation, and more loosely as a non-canonical label invoked by later benchmark papers.

1. Terminological status in the LIBERO literature

The term’s ambiguity is itself a substantive feature of the literature. CF-VLA does not introduce a benchmark titled “LIBERO-CF”; instead, it reports results on the existing LIBERO benchmark and describes a “coarse-to-fine” action generation mechanism that establishes a strong low-NFE operating point on LIBERO (Du et al., 27 Apr 2026). In that sense, “LIBERO-CF” functions as a convenient shorthand for LIBERO evaluated under coarse-to-fine action generation rather than as the name of a formally separate benchmark.

Later papers make the ambiguity explicit. LIBERO-Safety states that earlier LIBERO variants such as “LIBERO (and variants like LIBERO-Plus/PRO/X/CF)” primarily focus on task completion, generalization, and robustness rather than explicit safety, but it also adds that “LIBERO-CF (Collision-Free) is not an established official LIBERO benchmark per the cited paper” and may instead refer to collision-free data subsets in prior works (Cui et al., 22 Jun 2026). LIBERO-Para, by contrast, places “LIBERO-CF” in a different conceptual slot, describing it as “Counterfactual Failures, Fang et al. 2026,” complementary to language-only paraphrase perturbations (Kim et al., 30 Mar 2026). These usages do not define a single canonical artifact.

2. Substrate benchmark: LIBERO

Any account of LIBERO-CF must begin with LIBERO itself. LIBERO is a benchmark for lifelong learning in robot manipulation, or Lifelong Learning in Decision Making, designed to study transfer of both declarative knowledge and procedural knowledge (Liu et al., 2023). It introduces an extendible procedural generation pipeline grounded in language templates, PDDL-based scene descriptions, predicate-based goals, and robosuite as the underlying simulator.

For benchmarking, LIBERO provides four task suites, 130 tasks in total: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-100, with LIBERO-100 further split into LIBERO-90 for pretraining and LIBERO-Long for evaluation (Liu et al., 2023). The benchmark also provides 50 high-quality, human-teleoperated demonstrations for each of the 130 tasks, enabling imitation learning via behavioral cloning. In CF-VLA’s LIBERO evaluation, training is conducted on replay-filtered LIBERO trajectories from four suites—Spatial, Object, Goal, and Long—using demonstrations with successful outcomes (Du et al., 27 Apr 2026).

This benchmark context matters because the coarse-to-fine contribution is not a new scene generator or a new taxonomy of perturbations. Rather, it is an action-generation mechanism evaluated within LIBERO’s established manipulation suites, where the primary metric is task success rate and the experimental emphasis is the efficiency–performance trade-off under low numbers of function evaluations (Du et al., 27 Apr 2026).

3. Coarse-to-fine action generation in CF-VLA

CF-VLA is a flow-based Vision-Language-Action policy that addresses the inefficiency of multi-step inference from uninformative Gaussian noise by changing the role of the starting point in generative action modeling (Du et al., 27 Apr 2026). Its backbone is a standard Vision-Language-Model, such as PaliGemma, which receives visual observations, language instructions, and proprioceptive state as input. The architectural novelty lies in the action expert, which is decomposed into a coarse stage and a fine stage.

The coarse stage predicts a distribution over endpoint velocity and transforms Gaussian noise into an action-prior-guided initialization close to the action manifold. Concretely, the process begins by sampling Gaussian noise $\epsilon_1 \sim \mathcal{N}(0, I)$, predicting a conditional posterior over endpoint velocity $q_\theta(u \mid o, \epsilon_1)$, and transforming the noise toward an AP-guided initialization via

$$\tilde{\epsilon} = \epsilon_1 - \hat{u}, \quad \hat{u} \sim q_\theta(u \mid o, \epsilon_1).$$

The fine stage then takes $\tilde{\epsilon}$ and applies a local correction in a single step to recover the final action (Du et al., 27 Apr 2026).

The paper contrasts this with standard flow-matching interpolation,

$$x_t = t\epsilon + (1-t)a;\qquad u_t = \epsilon - a,$$

where iterative denoising or refinement is typically required to recover action structure from noise (Du et al., 27 Apr 2026). The central claim is not merely that fewer iterations are used, but that a structured initialization can make low-NFE inference viable.

4. Training procedure and objective

CF-VLA uses a stepwise training strategy intended to stabilize the coupling between coarse initialization and fine refinement (Du et al., 27 Apr 2026). Training alternates between a stabilization phase, in which proxy inputs support reliable coarse output learning for the refinement branch, and a subsequent joint optimization phase in which coarse and fine stages are coupled.

The joint loss is given as

$$\mathcal{L}(\theta) = \mathcal{L}_{\text{fine}}(\theta) + \lambda\,\mathcal{L}_{\text{coarse}}(\theta),$$

where $\mathcal{L}_{\text{coarse}}$ is the KL divergence between predicted and target endpoint-velocity distributions, and $\mathcal{L}_{\text{fine}}$ is the MSE between predicted and true local corrections (Du et al., 27 Apr 2026).

On LIBERO, the model is trained for 60,000 steps using 8×A100 GPUs, batch size 16, and learning rate $5 \times 10^{-5}$, with the same recipe applied to compared baselines including PI_0.5 and MIP variants (Du et al., 27 Apr 2026). Evaluation is performed with 10 simulation rollouts per task, measuring success rate per LIBERO suite and then averaging. The main reported metrics are success rate, action sampling latency excluding visual-language encoding and KV-cache, and Number of Function Evaluations.

5. Reported LIBERO results

On LIBERO simulation, CF-VLA reports 96.5% average success at NFE = 2, with suite-wise scores of 98.0 on Spatial, 99.2 on Object, 96.6 on Goal, and 92.0 on Long (Du et al., 27 Apr 2026). Under the same evaluation, MIP $(\pi_0)$ at NFE = 2 reports 92.7 average success; MIP $q_\theta(u \mid o, \epsilon_1)$0 at NFE = 2 reports 93.6; $q_\theta(u \mid o, \epsilon_1)$1 at NFE = 2 reports 94.8; and $q_\theta(u \mid o, \epsilon_1)$2 at NFE = 10 reports 95.7 (Du et al., 27 Apr 2026).

The paper emphasizes that CF-VLA at NFE = 2 matches or beats the NFE = 10 $q_\theta(u \mid o, \epsilon_1)$3 baseline on several metrics, while reducing action sampling latency from 29.17 ms to 7.81 ms, a 75.4% reduction (Du et al., 27 Apr 2026). It also reports especially strong gains on the Object and Long suites, described as tasks sensitive to long-term consistency and structured generation. Ablation studies are summarized as showing that removing phase-II coupling or the fine refinement step significantly degrades performance, indicating that both stages are necessary for the reported gains.

The same work reports an average real-robot success rate of 83.0%, outperforming MIP by 19.5 points and $q_\theta(u \mid o, \epsilon_1)$4 by 4.0 points (Du et al., 27 Apr 2026). The cited challenging tasks include Wipe Table, Pour Water, and Fold Towel into Thirds, which the paper characterizes as requiring long-horizon consistency and precise physical control.

6. Significance and relation to adjacent LIBERO variants

Within the CF-VLA paper, the significance of the LIBERO results is framed as an efficiency–performance frontier: the key design move is restructuring the starting point of action generation through AP-guided initialization rather than relying on more denoising or refinement steps (Du et al., 27 Apr 2026). The method is described as plug-and-play for arbitrary flow-based VLA policies and different VLM backbones with minimal modification. This suggests that the most consequential aspect of “LIBERO-CF” in this sense is methodological rather than taxonomic: it is a way of operating on LIBERO that tests whether structured initialization can displace iterative inference.

That emphasis differs from later LIBERO-family extensions, each of which isolates a distinct failure mode of VLA evaluation. LIBERO-X introduces a five-level hierarchical evaluation protocol for spatial generalization, object recognition, and task instruction understanding under cumulative perturbations (Wang et al., 6 Feb 2026). LIBERO-Occ studies scene-induced occlusion and proposes Viewpoint Imagination as a complementary-view generation mechanism (Li et al., 9 Jun 2026). LIBERO-Para isolates paraphrase robustness and introduces PRIDE as a difficulty-aware metric (Kim et al., 30 Mar 2026). LIBERO-PRO argues that standard LIBERO evaluation inflates performance estimates through memorization and proposes controlled perturbations over manipulated objects, initial states, instructions, and environments (Zhou et al., 4 Oct 2025). LIBERO-Safety adds explicit physical and semantic safety constraints, along with a keypose-driven collision-free data generation pipeline (Cui et al., 22 Jun 2026).

Against that background, “LIBERO-CF” does not denote a parallel benchmark family with its own formal protocol in the same way as LIBERO-X, LIBERO-Occ, LIBERO-Para, LIBERO-PRO, or LIBERO-Safety. Instead, the evidence in the cited literature indicates a context-dependent label attached either to coarse-to-fine generation on LIBERO, to collision-free subsets, or to counterfactual-failure evaluation, depending on the paper (Du et al., 27 Apr 2026, Cui et al., 22 Jun 2026, Kim et al., 30 Mar 2026). The most defensible encyclopedic characterization is therefore that LIBERO-CF is a non-canonical term whose most concrete instantiation is CF-VLA’s coarse-to-fine action-generation study on the LIBERO benchmark.