Tactile Modality Fusion for Vision-Language-Action Models¶

Conference: ECCV 2026
Paper: ECCV 2026 | Project Page
Area: Robotics & Embodied AI
Keywords: vision-language-action model, tactile fusion, FiLM modulation, contact-rich manipulation, parameter-efficient fine-tuning

TL;DR¶

TacFiLM turns a pretrained tactile encoder's embedding into per-channel FiLM scale/shift parameters that modulate the intermediate visual features of OpenVLA-OFT — without adding a single token to the language model — raising the in-distribution average success rate from 58.10% to 86.67% on real-robot peg insertion, USB/HDMI plugging and drawer-opening tasks while cutting peak contact force to roughly one third.

Background & Motivation¶

Vision-language-action (VLA) models — OpenVLA, RT-2-X, π0.5, OpenVLA-OFT and others — ground the semantic representations learned by vision-language models into low-level action generation, giving robot policies both semantic generalization and direct control output. Their perception, however, is almost exclusively visual. Contact-rich manipulation is precisely where vision is least useful: targets are occluded by the fingers and the workpiece itself, assembly tolerances are often at the millimetre scale, and the physical quantities that actually decide success — contact force, surface friction, material compliance, tangential shear — are simply not in the image. Humans use fingertip touch for these fine adjustments; robots have to guess from pixels. Vision-based tactile sensors such as DIGIT and GelSight render gel deformation as images, which makes them structurally compatible with vision architectures, and over the past two years a line of work has tried to push touch into VLA models.

Those attempts split into two routes, each with its own price. The first introduces touch at post-training finetuning time: tactile signals are encoded into additional tokens and concatenated onto the VLM input, or dedicated cross-attention layers are stacked after the vision backbone so that visual and tactile representations attend to each other. Concatenation lengthens the sequence and raises compute, and prior work has shown that performance can actually degrade as context grows; cross-attention adds a set of new trainable attention parameters that need more data to learn effective cross-modal correspondences. The second route relies on large-scale multimodal pretraining or contrastive learning to align touch with vision before policy learning — effective, but with a high data and compute barrier. Real deployments face a different reality: behaviour models are already huge, and a new task or a new sensor typically comes with only tens to a few hundred demonstrations. What is actually needed is a lightweight fusion scheme that fits inside the post-training finetuning paradigm, does not retrain large components, and still genuinely exploits tactile information.

The angle taken here is to stop inserting new tokens into the sequence and instead use touch to modulate visual features that already exist. Tactile images are encoded by a pretrained tactile encoder into a pooled embedding, a lightweight MLP turns it into FiLM (feature-wise linear modulation) per-channel scale and shift parameters, and those are applied inside the visual ViT blocks — after normalization, before multi-head self-attention. The language model sees exactly the same tokens as before, and no per-sensor encoder has to be trained. Core idea: recast tactile fusion from "concatenation / cross-attention" into "per-channel affine modulation (FiLM) of intermediate visual features", with zero-initialized \(\gamma,\beta\) so that conditioning starts from a near-identity mapping — injecting contact information without destroying the VLA's pretrained vision-language priors and without increasing token length.

Method¶

Overall Architecture¶

TacFiLM builds on OpenVLA-OFT: a fused SigLIP + DINOv2 vision backbone produces patch-level visual embeddings, a lightweight MLP projector maps them into the language model's input space, the decoder is Llama2 7B, and its final hidden states go to an MLP action head that regresses continuous action chunks with an L1 objective. TacFiLM inserts one conditioning stage inside the vision backbone: at each time step, DIGIT tactile images are encoded by a pretrained tactile encoder and pooled into a global tactile vector \(z_t\), which an MLP maps to a pair \(\gamma_n,\beta_n\) for each selected ViT block \(n\); those parameters apply a feature-wise affine transform to the intermediate visual features after normalization and before multi-head self-attention. The modulated visual features then proceed as usual: concatenated with language tokens and fed to the language model to produce action chunks.

Along this entire path, touch neither adds tokens nor adds attention layers — its only interface is the pair \(\gamma,\beta\) inside the ViT blocks. On the training side, LoRA adapts the linear layers of the VLA and the tactile backbone while the FiLM layers are trained from scratch; everything else stays frozen.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["DIGIT tactile frames + RealSense RGB<br/>+ language prompt (10 Hz)"] --> B["Pretrained tactile representations<br/>Sparsh-DINO encode → pooled z_t"]
    A --> C["SigLIP + DINOv2 vision backbone<br/>patch-level visual embeddings"]
    B --> D["FiLM tactile conditioning<br/>MLP emits γ/β, modulates features"]
    C --> D
    D --> E["Llama2 7B decoder<br/>visual tokens + language tokens"]
    E --> F["MLP action head · L1 regression"]
    F --> G["7-DoF continuous action chunks"]
    subgraph S["Parameter-efficient post-training (training side)"]
        direction TB
        H["LoRA: VLA and tactile backbone linear layers"] --> I["FiLM layers trained from scratch<br/>γ/β zero-initialized"]
    end
    I -.->|updates γ/β at training time| D

Key Designs¶

1. Reusing pretrained tactile representations: no task-specific encoder, just plug in an existing tactile backbone

In both prior fusion routes the encoder is either trained jointly with the policy or aligned through additional multimodal pretraining, and either option costs more data and compute. TacFiLM treats the tactile encoder as an off-the-shelf, frozen source of representations: tactile images get only minimal preprocessing (resize to 224×224, background removal; the Sparsh variants additionally concatenate two frames five time steps apart channel-wise to carry explicit dynamics), and their encoded features are used directly. Two architectures are evaluated. T3 is a framework with an independent ViT encoder per sensor plus a shared ViT transformer trunk; TacFiLM keeps only the sensor encoder and the trunk and discards the task-specific decoders. Sparsh is a ViT trained with self-supervised objectives, evaluated in all three proposed variants — Sparsh-MAE (masked autoencoding), Sparsh-IJEPA (joint-embedding prediction) and Sparsh-DINO (self-distillation). Both architectures produce fixed-dimensional embeddings, so the fusion module is agnostic to the encoder choice.

That choice is made empirically rather than by assumption: the pretrained representations are scored on three binary classification probes — Rotation-High / Rotation-Low, which ask whether the peg in the gripper has been rotated relative to its initial pose (sensitivity to changes in contact geometry), and Contact, which asks whether contact has occurred (encoding of force and deformation features) — plus the continuous force-regression task from the Sparsh TacBench benchmark, measured by RMSE. Sparsh-DINO is best on all three classification tasks with a 97.72% average and the lowest force RMSE at 36.09, so it becomes the tactile backbone for every later experiment. The payoff of this design is that changing sensors only means swapping in another off-the-shelf pretrained encoder, with no sensor-specific data collection or encoder training — which is exactly the "design philosophy of generalist VLAs" the paper appeals to.

2. FiLM tactile conditioning: modulate intermediate visual features instead of appending tokens to the input sequence

Concatenation-based methods project tactile embeddings and append them to visual/language tokens, lengthening the sequence and raising compute; cross-attention-based methods stack attention blocks after the vision backbone, and their extra trainable parameters need more data to converge. TacFiLM takes a third path and makes touch a global, low-dimensional bias acting on visual features. At each time step \(t\), the tactile image is encoded into patch features and average-pooled into \(z_t\); an MLP then produces a pair of channel-wise parameters \(\gamma_n,\beta_n\) for each selected ViT block \(n\) of both vision encoders (DINOv2 and SigLIP), applied to the intermediate visual features \(F_n\) after normalization and before multi-head self-attention:

\[\text{FiLM}(F_n \mid \gamma_n, \beta_n) = F_n \odot \gamma_n + \beta_n\]

(⚠️ The cached full text has this equation corrupted by PDF extraction into FiLM(F_n|γ_n,β_n)F_n⊙γ_n)β_n; it is restored here as the expression above from the original FiLM formulation and the surrounding context — refer to the original paper.) Following the design principles of OpenVLA-OFT, \(\gamma\) and \(\beta\) are applied to the entire feature map, and they are zero-initialized, so modulation starts as an identity mapping and the tactile pathway is grown during finetuning — which is why it does not disturb the VLA's vision-language priors from the first step. By default every ViT block is conditioned; Section 4.3 ablates this choice.

The design is lightweight because tactile information is compressed into a global modulation signal that is independent of sequence length: the language model sees exactly the same number of tokens as the vision-language baseline, so there is no context-bloat risk. Placing the modulation inside the ViT blocks also means touch can reshape visual representations before the language model reads them, rather than relying on attention to fuse them after the tokens reach the LLM. The ablation supports the "no need to cover everything" reading as well: conditioning only a subset of blocks barely changes performance.

3. Parameter-efficient post-training: learn only FiLM and a few LoRA weights, injecting touch rather than rebuilding the model

Even after large-scale pretraining, a generalist VLA is not accurate enough for specialized assembly tasks, but full finetuning is both expensive and prone to washing out generalist priors. TacFiLM follows the LoRA-style efficient finetuning recipe and extends it to the tactile dimension: LoRA adapters are attached to the linear layers of the TacFiLM-augmented VLA (both OpenVLA-OFT and the tactile backbone), while the FiLM layers are trained from scratch as new modules and the rest of the model stays frozen. The newly learnable parameters are therefore only two groups — the low-rank adapters and the \(\gamma,\beta\) emitted by those MLPs — yet they suffice to attach an entirely new modality to the existing policy.

This arrangement pays off in a way that echoes design 2: zero-initialized FiLM makes the tactile pathway start as an identity mapping that changes no behaviour, and LoRA only allows small adjustments to the base model, so the finetuned policy still stands on the original VLA's semantic understanding and merely gains sensitivity to contact state. The paper reports training all deployed methods to 80k steps.

A Worked Example¶

Take the easiest in-distribution task, Circle-Peg with 3mm clearance, and walk one control step (the flow and magnitudes come from the paper's task setup; internal tensor dimensions are not given in the original): the robot is set up so that the peg is already in hand — the paper explicitly does not evaluate grasping — and the current prompt is the templated "Insert the [colour] [shape] peg into the [colour] base". In one 10 Hz step, the RealSense RGB frame goes through SigLIP + DINOv2 to patch-level visual embeddings, the DIGIT sensor captures the gel deformation at the fingertip, Sparsh-DINO encodes it and pools it into a tactile vector \(z_t\), and the MLP immediately emits the \(\gamma_t,\beta_t\) for that step to modulate the visual features. If the peg's edge has only lightly grazed one side of the hole (nearly invisible in the image, since the peg and the gripper hide most of the opening), the deformation pattern is clearly one-sided, the modulation shifts accordingly, and the emitted action chunk is a small lateral correction. Pushing ahead on the vision-only action instead drives contact force up to the 14.94 N range — exactly what the vision-only baseline does at this clearance — whereas TacFiLM keeps this task at a 7.64 N peak force, 52.03 s average completion and a 36.67% first-attempt insertion rate. The same pipeline repeats on 2mm clearances and on the out-of-distribution square/pentagon pegs and HDMI plug; only the shape word in the prompt and the contact pattern on the sensor change.

Loss & Training¶

The action head regresses continuous action chunks directly with an L1 objective: the decoder's final hidden states pass through an MLP action head (the paper does not report the chunk length or action dimensionality). The optimization strategy is "frozen by default, locally learnable": LoRA is applied only to the linear layers of the TacFiLM-augmented VLA (covering both OpenVLA-OFT and the tactile backbone), the FiLM layers are trained from scratch with zero-initialized \(\gamma,\beta\), and all compared methods are trained to 80k steps for a fair comparison.

For data, the authors use Polymetis for high-level control and teleoperation and libfranka over the Franka Control Interface to send low-level commands at 1 kHz, with an operator driving the end-effector pose through a 3Dconnexion SpaceMouse whose 6-DoF input maps to Cartesian pose commands. Each task contributes 80 expert demonstrations of roughly 70 steps, recorded at 10 Hz with time-aligned joint positions and velocities, end-effector pose, gripper width and status, RGB images from an Intel RealSense camera, tactile images from a DIGIT sensor, and the executed actions, each accompanied by a fixed natural language description. The hardware is a 7-DoF Franka Emika Panda with the parallel two-finger Franka hand carrying a DIGIT visuotactile sensor.

Key Experimental Results¶

Main Results¶

All evaluation is on real hardware, totalling over 1,000 rollouts: 480 in-distribution (30 per method per task), 300 out-of-distribution (15 per method) and 240 for ablations. The in-distribution tasks are Circle-Peg insertion (3mm and 2mm clearance), USB cable plugging and drawer opening; the out-of-distribution tasks are Square-Peg and Pentagon-Peg insertion (3mm and 2mm clearance each) plus HDMI cable plugging. All four baselines share the same OpenVLA-OFT backbone so that differences are attributable to the fusion strategy: OpenVLA-OFT is the vision-only baseline; TactileConcat projects pretrained tactile embeddings through a two-layer MLP into the VLM input space and concatenates them with the visual and language tokens; Cross-Attn follows PolyTouch's architecture with six residual cross-attention blocks after the vision backbone, where visual patch embeddings serve as queries and tactile embeddings as keys/values.

Task	Method	Success (%)	Direct (%)	Avg. Max Force (N)	Avg. Time (s)
In-distribution (30 rollouts per method per task)
Circle-Peg 3mm	OpenVLA-OFT	86.67	3.33	14.94 ± 4.66	92.24 ± 48.10
	TactileConcat	96.67	16.67	9.19 ± 3.45	75.11 ± 37.28
	Cross-Attn	63.33	10.00	8.16 ± 7.18	155.13 ± 31.59
	TacFiLM	100.00	36.67	7.64 ± 2.63	52.03 ± 5.02
USB-Cable-Plug	OpenVLA-OFT	33.33	0.00	15.01 ± 9.09	164.52 ± 27.06
	TactileConcat	43.33	6.67	12.96 ± 5.04	135.11 ± 56.82
	Cross-Attn	33.33	0.00	26.23 ± 14.98	134.58 ± 52.96
	TacFiLM	73.33	33.33	10.15 ± 5.47	99.71 ± 46.43
Open-Drawer	OpenVLA-OFT	33.33	33.33	13.64 ± 0.54	152.65 ± 39.37
	TactileConcat	26.67	26.67	9.74 ± 0.89	141.66 ± 35.93
	Cross-Attn	20.00	20.00	17.40 ± 6.58	165.64 ± 28.61
	TacFiLM	86.67	73.33	10.84 ± 1.87	94.33 ± 46.43
ID average (as reported)	OpenVLA-OFT	58.10	12.38	14.94 ± 9.16	126.72 ± 51.34
	TactileConcat	64.76	10.48	10.27 ± 4.12	113.04 ± 52.31
	Cross-Attn	48.00	12.00	13.43 ± 12.62	149.92 ± 39.01
	TacFiLM	86.67	37.14	8.65 ± 3.80	81.72 ± 38.00
Out-of-distribution (15 rollouts per method per task)
Square-Peg 3mm	OpenVLA-OFT	93.33	0.00	18.31 ± 8.84	51.60 ± 5.62
	TactileConcat	93.33	13.33	9.34 ± 5.56	69.77 ± 23.40
	Cross-Attn	60.00	6.67	6.49 ± 1.14	165.27 ± 18.74
	TacFiLM	100.00	46.67	5.37 ± 0.41	52.95 ± 4.68
Square-Peg 2mm	OpenVLA-OFT	66.67	0.00	34.30 ± 11.18	64.54 ± 10.64
	TactileConcat	86.67	6.67	27.72 ± 8.36	91.83 ± 10.64
	Cross-Attn	53.33	6.67	26.24 ± 12.36	156.33 ± 28.07
	TacFiLM	80.00	40.00	7.06 ± 1.36	111.61 ± 38.31
HDMI-Cable-Plug	OpenVLA-OFT	6.67	0.00	10.71 ± 8.93	166.88 ± 38.29
	TactileConcat	13.33	0.00	11.18 ± 5.08	174.59 ± 13.84
	Cross-Attn	33.33	0.00	33.82 ± 12.79	133.97 ± 39.33
	TacFiLM	66.67	6.67	11.54 ± 3.88	116.87 ± 45.46
OOD average (as reported)	OpenVLA-OFT	54.67	0.00	22.46 ± 15.75	89.48 ± 46.05
	TactileConcat	73.33	8.00	16.47 ± 10.54	105.79 ± 43.16
	Cross-Attn	49.33	5.33	19.27 ± 14.62	149.77 ± 33.72
	TacFiLM	86.67	29.33	8.40 ± 4.71	87.84 ± 42.69

⚠️ The "average" rows are reproduced verbatim from the paper. The out-of-distribution averages are consistent with the per-task columns, but the in-distribution averages are not simple means of those columns (e.g. TacFiLM's four in-distribution direct-insertion rates are 36.67 / 23.33 / 33.33 / 73.33, a mean of 41.67%, whereas the paper reports 37.14%; the vision-only baseline's success column averages 55.00% versus the reported 58.10%). This is presumably rollout-weighted or a typesetting/extraction artefact. Statements in the prose follow the per-task values in the table above.

Ablation Study¶

The first ablation asks where FiLM should be inserted: the default conditions every ViT block (AllFiLM), and three variants condition only one third of the blocks at different depths — EarlyFiLM, MiddleFiLM and LateFiLM for the shallow, middle and deep layers of the vision backbone. The second examines whether touch can compensate when the camera degrades: lighting dimmed by 80%, and a partially frozen stream where only 50% of frames update.

Setting	Method	Success (%)	Direct (%)	Avg. Max Force (N)	Avg. Time (s)
Circle-Peg 3mm (ID)	AllFiLM	100.00	36.67	7.64 ± 2.63	52.03 ± 5.02
	EarlyFiLM	93.33	60.00	6.69 ± 0.92	60.85 ± 8.26
	MiddleFiLM	100.00	26.67	7.14 ± 1.71	53.88 ± 5.08
	LateFiLM	100.00	23.33	8.47 ± 2.71	54.74 ± 7.34
Pentagon-Peg 3mm (OOD)	AllFiLM	100.00	33.33	7.51 ± 2.88	53.15 ± 5.89
	EarlyFiLM	100.00	33.33	9.96 ± 4.49	77.40 ± 21.68
	MiddleFiLM	100.00	53.33	8.99 ± 4.49	53.57 ± 6.89
	LateFiLM	100.00	40.00	9.42 ± 3.31	68.57 ± 27.47
Circle-Peg 3mm (80% dimmed)	OpenVLA-OFT	93.33	0.00	16.29 ± 9.85	73.50 ± 8.41
	TactileConcat	86.67	26.67	11.15 ± 9.21	78.03 ± 30.16
	Cross-Attn	53.33	0.00	9.29 ± 5.96	159.96 ± 29.15
	TacFiLM	100.00	26.67	8.62 ± 2.13	67.79 ± 6.25
Circle-Peg 3mm (50% frames)	OpenVLA-OFT	73.33	0.00	15.70 ± 12.39	113.19 ± 40.39
	TactileConcat	80.00	40.00	14.64 ± 13.91	71.52 ± 34.51
	Cross-Attn	46.67	0.00	7.53 ± 1.25	161.10 ± 30.70
	TacFiLM	100.00	26.67	8.12 ± 1.90	51.68 ± 5.33

A third study selects the tactile encoder, using the three binary classification probes plus the TacBench force-regression RMSE.

Task	T3	Sparsh-IJEPA	Sparsh-MAE	Sparsh-DINO
Rotation-High (%)	92.73	99.15	99.36	99.36
Rotation-Low (%)	83.09	96.44	96.64	98.42
Contact (%)	73.31	85.08	93.92	95.39
Classification average (%)	83.04	93.56	96.64	97.72
Force estimation RMSE (TacBench)	58.64	40.27	36.61	36.09

Key Findings¶

Tactile gains concentrate where vision is physically blind. The three largest gains are exactly the tasks where vision is most disadvantaged: drawer opening 33.33% → 86.67% (+53.3 points), in-distribution USB plugging 33.33% → 73.33%, and out-of-distribution HDMI plugging 6.67% / 13.33% → 66.67% (+33.3 points, described as "improves HDMI cable plugging success by 30%" in the abstract). The 2mm clearance settings behave the same way: vision cannot resolve millimetre-level offsets, so both success and first-attempt rates depend on touch.
The improvement shows up more in "getting it right the first time" than in recovery. In-distribution average success goes from 58.10% to 86.67%, while the direct-insertion rate goes from 12.38% to 37.14% (reported averages), meaning touch makes the policy align accurately on the first attempt instead of probing repeatedly. Drawer opening is the extreme case: TacFiLM reaches a 73.33% direct rate while no other method exceeds 33.33%.
Touch also solves force safety as a side effect. TacFiLM attains the lowest or near-lowest peak contact force on every task: 8.65 N versus 14.94 N for vision-only in distribution, and 8.40 N versus 22.46 N out of distribution — roughly one third of the baselines. TactileConcat, by contrast, jams on the hard settings: on the out-of-distribution Square-Peg 2mm it reaches 86.67% success but pushes peak force to 27.72 N, where TacFiLM needs only 7.06 N on the same task.
Cross-attention suffers most when data is scarce. Cross-Attn is last on every task, with a 48.00% in-distribution average success and 149.92 s average completion, and even on its relatively best task (HDMI) it trails TacFiLM. The authors attribute this to the additional trainable visual-tactile interactions it introduces: with only 80 demonstrations per task and 80k training steps, it cannot learn reliable multimodal correspondences, whereas FiLM compresses touch into a global modulation bias with far lower parameter demand.
FiLM does not need to cover every ViT block. The four integration locations are nearly tied on success (93.33%–100% on Circle-Peg 3mm, 100% across the board on Pentagon-Peg 3mm), but EarlyFiLM raises the direct-insertion rate on Circle-Peg 3mm from AllFiLM's 36.67% to 60.00%, suggesting shallow conditioning may already be sufficient and that a deployment could cut overhead further by touching only a few blocks.
Touch is a reliable backup channel under camera degradation. Under both 80% dimming and 50% frame updates, TacFiLM holds a 100% success rate, while the vision-only baseline falls to 73.33% with frozen frames and Cross-Attn drops from 63.33% to 46.67%–53.33%. Interestingly, TactileConcat has a higher direct rate (40.00%) with frozen frames, but pays for it with larger contact forces and slower execution — its tactile tokens do work, just not as well as FiLM at controlling force and time simultaneously.
Encoder choice: stronger self-supervision yields tactile features better suited to insertion. T3 reaches only 73.31% on the Contact probe and a 58.64 force RMSE, clearly behind the three Sparsh variants, and Sparsh-DINO is best across the board (97.72% classification average, 36.09 RMSE). This also confirms design 1's premise — the quality of the pretrained tactile representation directly bounds how far lightweight fusion can go.

Highlights & Insights¶

Moving fusion from sequence space into feature space. Both concatenation and cross-attention modify the language model's input or add new attention layers, whereas FiLM only rescales and shifts visual features channel-wise. That buys three things at once: the token count is unchanged (no long-context degradation risk), the parameter cost is tiny (one MLP emitting two vectors), and the module is decoupled from the sensor (swapping the tactile encoder leaves fusion untouched). The idea transfers to any setting where a large model must absorb a low-bandwidth sensor — force/torque, audio or temperature, for instance.
Zero-initialized conditioning is what protects the pretrained priors. Initializing \(\gamma,\beta\) to zero makes the training starting point equivalent to the original model, and the tactile pathway is grown from an identity mapping; this explains why finetuning on a task with only 80 demonstrations does not wash out the vision-language priors. The trick (conditioning starts near identity) can be moved into essentially any FiLM / AdaLN-style adaptation at no cost.
Picking the encoder with probes and regression rather than end-to-end trials. Instead of running full policy experiments with T3 and each Sparsh variant, the authors first screen them on two binary probes (rotation, contact) and TacBench force regression, then run the policy study with the winner. The same low-cost selection protocol applies to any work that bolts a pretrained representation onto a policy.
Early-layer conditioning may beat conditioning everything. EarlyFiLM reaches a 60.00% direct rate against AllFiLM's 36.67% on Circle-Peg 3mm (6.67 points lower on success), which suggests touch behaves more like a prior that should be injected early than a late correction signal — a useful hint for where to place tactile-vision fusion in future designs.

Limitations & Future Work¶

The limitation the authors admit first is task coverage. Without precise visuotactile simulators, everything had to run on real hardware, making data collection and rollouts extremely time-consuming, so the study covers only insertion and drawer pulling rather than a broader manipulation suite (dexterous in-hand manipulation, deformable objects).
The method is tied to the OpenVLA-OFT architecture: FiLM is inserted into its vision backbone blocks, and adapting it to other VLA backbones such as π0.5 is left as future work. Given how different VLA backbones differ in vision encoder and feature dimensionality, that transfer is not a one-line change.
From a reader's perspective there are further gaps. First, grasping is not evaluated — every insertion task starts with the object already in hand — which sidesteps one of touch's most valuable uses, slip detection and force-controlled grasping. Second, each task has only 80 demonstrations and a single object geometry, and the out-of-distribution setting merely swaps peg shape and clearance, so genuine cross-object or cross-sensor generalization is untested. Third, although the paper argues that reusing pretrained tactile representations removes the need for sensor-specific data collection, only the DIGIT sensor is used, so the cross-sensor claim is not backed by experiments. Fourth, the in-distribution average rows do not match the per-task columns, which does not change the trend but makes the precise numbers less trustworthy.
Plausible improvements: ablate modulation location jointly with whether touch needs temporal history (the two-frame channel stacking in Sparsh supplies dynamics implicitly today, whereas FiLM itself is per-frame); combine cross-attention with FiLM on hard settings such as 2mm clearance, using the latter for a low-dimensional prior and the former for fine-grained alignment; and build visuotactile pairs in simulation to scale evaluation beyond two task families.

vs TactileConcat (concatenation-based fusion): it projects pretrained tactile embeddings through a two-layer MLP into tokens appended to the VLM input, whereas this paper compresses touch into per-channel modulation of visual features. The two are close on the easy task (96.67% vs 100% on Circle-Peg 3mm), but concatenation decays quickly as tasks get harder: 64.76% vs 86.67% in-distribution average, and peak force spikes at the 2mm setting (27.72 N vs 7.06 N). The root difference is that concatenation treats touch as another equal token while this paper treats it as a bias acting on vision.
vs Cross-Attn (PolyTouch-style cross-attention): it aligns vision and touch with six residual cross-attention blocks after the vision backbone; this paper argues such added trainable interactions cannot be learned from 80 demonstrations per task, and the results support that (48.00% in-distribution average, below the 58.10% vision-only baseline). The lesson: when data is limited, the parameter count of the fusion module is itself a variable worth trading off.
vs VLA-Touch / VTLA / ForceVLA and similar tactile-augmented VLAs: these also introduce touch during post-training, but mostly through token concatenation or additional pretraining alignment; the difference here is the insistence on not increasing sequence length, not training task-specific encoders, and staying inside the parameter-efficient finetuning paradigm.
vs T3 / Sparsh (tactile representation pretraining): this paper does not improve the encoders themselves but uses their representations as plug-and-play inputs and contributes a low-cost encoder selection protocol (probe classification plus force regression). For multimodal policy builders the message is: pick the right representation source first, then decide how to fuse it.
vs vision-only VLAs (OpenVLA, OpenVLA-OFT, π0.5): the training paradigm of these models is untouched here; the paper only demonstrates how cost-effective adding a tactile channel is for contact-rich tasks — especially when vision is occluded, poorly lit or dropping frames, touch can hold the success rate on its own.

Rating¶

Novelty: ⭐⭐⭐⭐ [FiLM is not new in itself, but using it as the tactile fusion interface for a VLA, paired with a "pretrained tactile representation + zero-initialized modulation + LoRA" recipe, is a clear and previously rare combination]
Experimental Thoroughness: ⭐⭐⭐⭐⭐ [over 1,000 real-robot rollouts spanning four in-distribution and five out-of-distribution tasks, two clearance levels, four fusion methods, four FiLM placements and two camera-degradation conditions, plus a separate tactile-encoder comparison]
Writing Quality: ⭐⭐⭐ [motivation and design are stated clearly, but the in-distribution average rows do not match the per-task columns and some equations are corrupted in typesetting/extraction, leaving readers to verify numbers themselves]
Value: ⭐⭐⭐⭐ [offers a low-cost route to wiring touch into a VLA under limited demonstrations, with empirical evidence on when to use which fusion (avoid cross-attention when data is scarce)]