CRAFT: Force-Aware Curriculum Fine-Tuning

• CRAFT is a force-aware curriculum fine-tuning framework that prioritizes low-entropy force cues to enhance contact-rich manipulation.
• The framework integrates a Variational Information Bottleneck and torque-based proprioception to improve precision and safety in tasks like peg-in-hole and object rolling.
• Experimental results show significant improvements, with success rate gains of up to 35.4 points across various VLA architectures and manipulation tasks.

Searching arXiv for the specified paper and closely related VLA/contact-rich manipulation context. arxiv_search(query="(Zhang et al., 13 Feb 2026) CRAFT Adapting VLA Models to Contact-rich Manipulation via Force-aware Curriculum Fine-tuning", max_results=5) arxiv_search(query="Vision-Language-Action contact-rich manipulation force-aware curriculum fine-tuning", max_results=10) CRAFT is a force-aware curriculum fine-tuning framework for adapting pretrained Vision-Language-Action (VLA) models to contact-rich manipulation, a regime in which success depends on precise alignment, stable maintenance of contact force, and dynamic adaptation to deformable or compliant objects. Its central premise is that high-entropy vision and language inputs can dominate learning and drown out low-entropy but critical force cues, so the model should first be compelled to rely on force-related signals and only later regain full multimodal access. To support this training regime, CRAFT combines a variational information bottleneck, torque-based proprioception, and a homologous leader–follower teleoperation system that collects synchronized vision, language, and force data across diverse tasks (Zhang et al., 13 Feb 2026).

1. Problem formulation and motivation

Contact-rich manipulation includes tasks such as peg-in-hole insertions, deformable object rolling, and surface wiping. In these settings, the policy must achieve precise alignment of end-effector and target, maintain stable contact force to avoid slip or over-force, and adapt dynamically to deformable or compliant objects. The paper frames these requirements as a systematic challenge for general-purpose VLA policies, which are effective at mapping camera streams and natural-language goals to actions but remain vulnerable during contact-intensive phases (Zhang et al., 13 Feb 2026).

The stated failure mode is an imbalance between modalities. Visual and linguistic inputs are high-entropy, whereas force signals are low-entropy but critical. In the reported formulation, VLA models such as RT-1/RT-2, SayCan, RDT, and $\pi_0$ can therefore become over-reliant on perception, while sparse force cues are underused. Without explicit force feedback, policies misalign or apply unsafe forces in contact-intensive steps. CRAFT addresses this by enforcing a “force-first, then multimodal” curriculum: during early fine-tuning, vision and language richness is suppressed through an Information Bottleneck, compelling the policy to lean on joint-torque signals; later, the bottleneck is relaxed so that vision and language are reincorporated once force-sensitive behaviors have been learned (Zhang et al., 13 Feb 2026).

This suggests that the framework is not defined simply by adding another modality. Its defining claim is that modality scheduling matters: force is prioritized first, and multimodal fusion is restored only after a force-sensitive control prior has formed.

2. Force-aware curriculum fine-tuning

The core algorithmic component is a Variational Information Bottleneck (VIB) applied to pretrained vision and language embeddings. Let $F_V \in \mathbb{R}^d$ and $F_L \in \mathbb{R}^d$ denote the outputs of the vision and language encoders. Each is passed through a Gaussian VIB layer that produces compact latents $F_V^c$ and $F_L^c$ according to

$$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$

The bottleneck penalty upper-bounds the mutual information $I(F;F^c)$ through

$$L_{\mathrm{VIB}}(F) = D_{\mathrm{KL}}\!\left[p(F^c \mid F)\ \|\ \mathcal{N}(0,I)\right].$$

During training, the compact latent is sampled as

$$F^c = \mu(F) + \sigma(F)\odot \epsilon,\qquad \epsilon \sim \mathcal{N}(0,I).$$

The task objective is an imitation loss over a horizon $H$,

$F_V \in \mathbb{R}^d$0

and the total optimization target is

$F_V \in \mathbb{R}^d$1

The curriculum is encoded by an exponential gate,

$F_V \in \mathbb{R}^d$2

so that $F_V \in \mathbb{R}^d$3 is large at early training steps and decays toward zero later (Zhang et al., 13 Feb 2026).

The gating behavior is explicit. In early stages, $F_V \in \mathbb{R}^d$4 and vision-language embeddings are heavily regularized, so the policy effectively uses only force, represented by joint torques $F_V \in \mathbb{R}^d$5, plus a tiny fraction of visual-linguistic context. In mid stages, the penalty decays and vision/language begin to re-enter. In late stages, $F_V \in \mathbb{R}^d$6, and full multimodal fusion is restored. A plausible implication is that CRAFT treats contact competence as a curriculum-dependent representation problem rather than as a purely architectural fusion problem.

3. Adapted VLA architecture and force fusion

CRAFT is defined as an adaptation layer around an existing VLA backbone rather than a replacement for it. The base VLA architecture contains a visual encoder $F_V \in \mathbb{R}^d$7, a language encoder $F_V \in \mathbb{R}^d$8, proprioception from robot state, and a policy head $F_V \in \mathbb{R}^d$9 with flow or diffusion outputs. CRAFT adds two small modules and changes the proprioceptive signal (Zhang et al., 13 Feb 2026).

First, the VIB compresses $F_L \in \mathbb{R}^d$0 and $F_L \in \mathbb{R}^d$1 into $F_L \in \mathbb{R}^d$2 and $F_L \in \mathbb{R}^d$3. Second, raw joint positions $F_L \in \mathbb{R}^d$4 are replaced by joint torques $F_L \in \mathbb{R}^d$5 as proprioception. Under impedance control, the torque signal is written as

$F_L \in \mathbb{R}^d$6

In this formulation, $F_L \in \mathbb{R}^d$7 contains both kinematic errors and external contact force. The final fused input is

$F_L \in \mathbb{R}^d$8

which is passed through a lightweight fusion MLP and then into the flow or diffusion policy backbone (Zhang et al., 13 Feb 2026).

The reported rationale for replacing $F_L \in \mathbb{R}^d$9 with $F_V^c$0 is that torque-based proprioception encodes kinematic error, inertia, and external contact in a single low-dimensional signal. This is presented as a key mechanism for contact sensitivity. The paper also characterizes the design as plug-and-play: it adds only a VIB and a fusion MLP, and requires no architectural overhaul of existing VLA models.

4. Data collection system and training recipe

To enable force-aware learning, the framework includes a homologous leader–follower teleoperation system. The hardware uses two “homologous” Panda arms, one as a low-cost leader with internal force control and the other as the follower. During operation, the human operator moves the leader; the follower mirrors the commands under impedance control; and joint torques measured on the follower are sent back to the leader handle in real time, yielding bilateral synchronization (Zhang et al., 13 Feb 2026).

The perception stack uses two cameras: a wrist-mounted camera for egocentric observation and a static third-person camera for contextual observation. Each demonstration also includes a language annotation. The collected data are synchronized tuples of the form $F_V^c$1, with 50 demonstrations per task and randomized poses.

The fine-tuning dataset consists of 50 teleoperation demonstrations per task across five tasks. The input pipeline is $F_V^c$2 fusion $F_V^c$3 policy $F_V^c$4 actions. The loss is

$F_V^c$5

where $F_V^c$6 is an $F_V^c$7 loss between predicted and demonstration action sequences. The reported hyperparameters are $F_V^c$8, $F_V^c$9 steps, Adam with learning rate $F_L^c$0, batch size $F_L^c$1, and total fine-tuning length of approximately $F_L^c$2 steps. The pretrained backbones are $F_L^c$3 or RDT trained on vision and language (Zhang et al., 13 Feb 2026).

These implementation details matter because the claimed improvements are obtained under a relatively small teleoperation dataset. A plausible implication is that the curriculum is intended to extract more contact-relevant supervision from limited real-world data rather than to rely on large-scale force-labeled corpora.

5. Experimental evaluation

The real-world evaluation covers five tasks: USB insertion, shaft-to-hole insertion, flip carton, wipe whiteboard, and roll plasticine. The evaluation protocol uses 20 trials for insertion and carton tasks, and 12 trials for wiping and rolling. Baselines include $F_L^c$4-base and RDT-base without force, as well as FACTR, described as a Gaussian-blur curriculum plus force (Zhang et al., 13 Feb 2026).

The main quantitative result is an increase in average success rate across the five tasks for both tested backbones.

Model	w/o CRAFT	w/ CRAFT
$F_L^c$5-base	25.3%	60.7%
RDT	22.7%	48.3%

For $F_L^c$6-base, the improvement is $F_L^c$7 points; for RDT, it is $F_L^c$8 points. Per-task highlights include wipe whiteboard, where $F_L^c$9 CRAFT reaches $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$0 versus $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$1 for $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$2-base, and roll plasticine, where RDT + CRAFT reaches $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$3 versus $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$4 for RDT-base (Zhang et al., 13 Feb 2026).

The reported out-of-distribution generalization results include different carton shape and color and changed socket height and angle. Under these variants, $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$5-base averages $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$6, while $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$7 CRAFT reaches $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$8, a gain of $$p(F^c \mid F) = \mathcal{N}(\mu(F), \operatorname{diag}(\sigma^2(F))).$$9 points. The paper states that CRAFT consistently improves task success, generalizes to unseen objects and novel task variations, and adapts effectively across diverse VLA architectures.

The empirical significance is therefore twofold. First, the gains are not confined to a single architecture. Second, the improvements are reported not only on nominal task instances but also on object-level and task-level variants, which the paper uses to support the claim of robust and generalizable contact-rich manipulation.

6. Ablations, interpretation, and scope

The ablation study is conducted on $I(F;F^c)$0-base over USB insertion, wiping, and rolling, with 10 trials per task. Three variants are reported: no CRAFT, VIB with positions and no torque, and full CRAFT with VIB plus torque. Their average success rates are $I(F;F^c)$1, $I(F;F^c)$2, and $I(F;F^c)$3, respectively (Zhang et al., 13 Feb 2026).

These ablations support two distinct claims. First, VIB alone forces the model to use proprioceptive cues, improving success from $I(F;F^c)$4 to $I(F;F^c)$5. Second, replacing $I(F;F^c)$6 with $I(F;F^c)$7 further raises performance to $I(F;F^c)$8, which the paper interprets as confirmation that torque carries richer contact information. A common misconception would be to treat CRAFT as merely force augmentation; the ablation evidence instead separates the effect of the information bottleneck from the effect of torque-based proprioception.

The paper attributes CRAFT’s performance to three mechanisms. It prevents vision/language over-dominance by forcing the policy to internalize low-entropy force or tactile cues via an information-theoretic bottleneck; it uses torque-based proprioception to encode kinematic error, inertia, and external contact in a single low-dimensional signal; and it uses an exponential curriculum to re-integrate high-entropy multimodal cues after the “force foundation” is laid (Zhang et al., 13 Feb 2026).

The design scope is correspondingly narrow and specific. CRAFT is presented as a method for adapting pretrained VLA policies to precise, contact-rich manipulation through force-centered proprioception and curriculum gating. This suggests that its principal contribution lies not in redefining VLA modeling in general, but in specifying how a pretrained multimodal policy can be reconditioned for manipulation regimes where contact dynamics are decisive.