HapticWAM:Distilling Imagined Touch into a World–Action Model without Inference-Time Tactile Sensing

Learning to predict contact, then transferring that knowledge to a policy without fingertip observations.

Mikhail Sannikov*, Ilya Mikhalchuk*, Konstantin Gubernatorov*,
Petr Kovalev, Ogunwoye Faith Oluwatobi, Dzmitry Tsetserukou

Intelligent Space Robotics Laboratory · Skoltech * Equal contribution

01 / WafflesStudent · 1×
02 / CartonStudent · 1×
03 / EggStudent · 1×
Figure 1 Three recorded student trials, one per task, at real-time speed. Each is seed 101 and has an operator verdict of successful placement. These examples illustrate behavior; aggregate results appear below. Episode records
01 / Overview

Contact is more than appearance.

A scene image can show a gripper surrounding an object without resolving whether the grasp is stable, slipping, or applying excessive force. Optical tactile sensors expose local deformation and contact mechanics, but directly conditioning a policy on these measurements ties its inputs to the fingertips.

We introduce HapticWAM, a world–action model that jointly generates actions and a structured description of future contact. A tactile-conditioned teacher learns from gel images, mechanical fields, and contact state. Haptic-Imagination Distillation transfers its contact futures and action predictions to a student that retains contact generation while removing the fingertip input branches.

The central question is whether contact can remain a useful internal representation when it is no longer directly observed by the policy.

UR3 robot with a Robotiq 2F-85 gripper, paired DM-Tac W2L tactile fingertips, RealSense camera, and Echo teleoperation setup
Figure 2 The physical setup. A UR3 manipulator carries two optical tactile fingertips. Scene images, robot state, and fingertip streams are recorded together.
02 / Method

Predicting touch alongside action.

A frozen Cosmos-Predict2.5-2B video backbone processes observation, video, contact, and action groups in one latent sequence. LoRA adapters and newly introduced modules learn the manipulation task. At each replan, the model jointly denoises future video, a contact package, and an action chunk.

The contact package represents events and mechanics: no contact, onset, hold, slip, or release; changes in displacement and normal-force fields; contact masks, centers of pressure, slip scores, and wrenches. It does not merely reconstruct the appearance of a tactile image.

Framework overview

Full HapticWAM framework from the paper: text, scene RGB and gel images, FiLM-conditioned mechanical fields, robot state and wrist history enter a joint backbone; video, contact and action tokens feed their respective decoders, with event and uncertainty heads on contact states.
Figure 3a Framework overview, reproduced from the manuscript at its original resolution. Solid tokens denote clean conditioning; hatched tokens denote noisy generated groups. The dashed fingertip region contains the teacher-only observations. Flame symbols identify trainable modules; the video backbone is adapted with LoRA while its base weights remain frozen.

From sensor streams to latent frames

Appearance. The frozen VAE encodes the scene image and the two gel images. A learned pointwise convolution fuses the fingertip appearance into one observation frame. Scene RGB remains available to both teacher and student.

Mechanics. Each fingertip contributes eight field channels: two deformation components, depth, two shear components, and three distributed-force components. The 11-dimensional contact state contains the resultant wrench, contact area, center of pressure, slip, and mask fraction. A contact-state MLP produces FiLM modulation for the convolutional field encoder.

Robot context. A state MLP encodes the 26-dimensional robot state, and a causal temporal convolution encodes the arm-side wrench history. This wrench is estimated from motor currents. Task text and the previous action chunk provide additional conditioning.

Joint generation and readout

Observation frames remain fixed while video, contact, and action groups are jointly denoised using rectified flow. The VAE decodes video futures; the contact decoder unpacks spatial fields and mechanical quantities; and the action decoder returns end-effector pose deltas and absolute gripper apertures. Separate event and uncertainty heads read the contact hidden states.

The contact codec packs field changes and masks onto a latent grid, event labels into one-hot bands, centers of pressure into Gaussian bumps, slip into tiled values, and wrenches into spatial blocks. The event and mechanical channels are generated together. A structural attention mask blocks contact/action queries from directly attending to generated-video keys, although indirect coupling remains possible.

Compare teacher and student inputs
Observation → joint generation
Conditioning
Scene image It
Robot state + arm wrench
Task text + previous action
Fingertip observations Not an input to the student
World–action model

Frozen video backbone

Cosmos-Predict2.5-2B
+ LoRA + learned modules

Anticipatory Contact CouplingPrevious imagined contact → attention
Jointly generated
Video futures 3 s horizon
Contact package +1, +2, +3 s
Action chunk 16 steps · 1.6 s
Student · 12 latent frames · anticipatory branch only · 26.4 M trainable parameters

A compact view of the input difference. The detailed framework above and internals below specify the full architecture.

Inside the backbone: attention and distillation

The backbone contains 28 diffusion-transformer blocks, each with self-attention, cross-attention to the task-text embedding, and a feed-forward MLP. ACC modifies self-attention scores on haptic keys. HID supplies the training signals that transfer contact generation and action behavior from the frozen teacher to the student.

Framework internals

Full internal architecture from the paper: 28 DiT blocks; ACC fusion of wrench, visual tokens, previous action and imagined contact into gates, event distributions and typed attention bias; HID teacher and student branches connected by contact-future, event and behavior losses plus ground-truth supervision.
Figure 3b Backbone, Anticipatory Contact Coupling (ACC), and Haptic-Imagination Distillation (HID), shown in the updated author-supplied architecture figure. During HID, the teacher is frozen. Only contact-future matching uses the event/uncertainty weights. The teacher’s reactive gate is driven by observed tactile change; the student retains only the anticipatory branch.

Anticipatory Contact Coupling

ACC combines current observations with contact imagined at the previous replan, using this summary to bias attention on haptic keys. The teacher also has a reactive branch driven by observed tactile change. The student uses the anticipatory branch alone. At deployment, the prior contact package is generated by the model, rather than supplied by future measurements.

The fusion MLP combines the encoded wrench history, pooled conditioning-frame features, the previous action chunk, and the aligned previous contact prediction. It outputs an anticipatory gate, a future-event distribution, and a blending coefficient. The teacher mixes anticipatory and reactive gates; the student uses the anticipatory gate directly.

Teachergt = αt gtant + (1 − αt) gtreact
Studentgt = gtant

Each transformer block adds a learned, event-typed bias multiplied by this gate to pre-softmax scores on haptic keys. Block scales start at zero. This internal attention modulation is distinct from the scripted contact-probability veto used by the deployment controller.

Transferring contact knowledge

Haptic-Imagination Distillation matches the teacher’s contact futures, event distributions, and action velocities. Event saliency and predicted uncertainty weight contact-future matching. The student inherits teacher weights and is distilled on demonstrations and mixed policy rollouts, including unsuccessful rollouts with no ground-truth action weight.

Contact futures · Ltraj

Match the packed student contact prediction to the frozen teacher’s generated package. Per-step weights increase with predicted contact-event saliency and decrease with teacher-predicted uncertainty; this is not assumed to be calibrated confidence.

Events · LevtS

Match teacher and student event distributions using KL divergence. This loss does not receive the contact-future weighting.

Behavior · Lbeh

Match action-flow velocities at the same noisy action chunk and noise level. Contact and behavior targets come from the frozen teacher.

Additional supervision

Retain reduced-weight ground-truth action, video, and arm-wrench losses, ACC auxiliaries, and an uncertainty readout trained on detached hidden states. Failed rollouts have zero ground-truth action weight while teacher matching remains active.

Training configuration and physical horizons

The teacher has 27.9 M trainable parameters and receives 20,000 demonstration updates followed by 2,000 simulation/rollout updates. The deployed student receives 1,000 distillation updates. Removing its fused gel and mechanics observation frames reduces the latent sequence from 14 to 12 frames; the student has 26.4 M trainable parameters.

Each training clip contains one current RGB frame and 12 future targets at 4 Hz. Contact targets are at 1, 2, and 3 seconds; action targets are at 0.1–1.6 seconds at 10 Hz. Future targets are not observed history.

Full equations and loss weights in the manuscript
03 / Inspect a rollout

See the motion. Inspect the contact.

Follow a recorded execution through the scene camera, left/right tactile deformation heatmaps and their normal-force traces. The camera and heatmaps share one playback clock; the plot cursor follows the same recorded time.

Switch between selected successful teacher and student executions of each task. Starting seeds may differ; playback resets when you switch.

Synchronized rollout viewer

Loading recorded rollouts…

Figure 4 Selected successful executions. Starting seeds are shown for each recording. Tactile heatmaps show the absolute recorded depth field, not predicted contact masks. Camera frames and tactile fields are resampled by nearest timestamp to a single 15 fps video, without interpolated frames; force samples retain their original timestamps. These examples illustrate execution, not a controlled comparison of model inputs. Source episode
Recording and synchronization details

Fz is the recorded tactile wrench’s normal-force component, exported in newtons by the sensor driver. This is not the uncalibrated distributed force-change field used in the contact package. No independent calibration-accuracy claim is made.

Download timestamps, force traces and provenance
04 / Physical evaluation

Three tasks. Two hundred trials.

We evaluate the deployed teacher, distilled student, π0.5, and Diffusion Policy on waffle, carton, and egg pick-and-place. Each model is tested on 20 waffle, 20 carton, and 10 egg starts. Starting positions are shared across models; success is an operator-reviewed placement at the target.

The student achieves the highest observed placement rate on each task: 19/20 on waffles, 17/20 on cartons, and 5/10 on eggs. Across tasks, this is 41/50 placements (82% pooled), or 76.7% when each task is weighted equally.

Placement success
Figure 5 Points show placement rates; lines show descriptive 95% Wilson intervals. Counts are operator verdicts from the main comparison. Intervals describe binomial uncertainty, not variation across training seeds.
Exact counts and downloadable results
Successful placements / evaluated starts
ModelWafflesCartonEggPooled
Teacher14/2012/204/1030/50
Student19/2017/205/1041/50
π0.54/206/200/1010/50
Diffusion Policy2/203/204/109/50

Download summary CSV

These are comparisons of complete deployed configurations. Training data, candidate selection, gripper timing, and replan limits differ. The student–teacher gap is not separable at this sample size, and the comparison does not isolate a distillation effect. Egg handling remains difficult: its 5/10 student result has a wide 23.7–76.3% Wilson interval.

Evaluation protocol and baseline configuration

Starts use seeds 101–120 for waffles and cartons and 101–110 for eggs. Model order alternates between launches. The operator is not blinded; the reported teacher–student pair was selected after its carton result was observed.

The baseline export contains 876 episodes and excludes deliberate failures; baselines receive no simulation episodes or policy rollouts. Diffusion Policy’s deployment history is 0.486 s apart, compared with 0.1 s during training.

The teacher uses compiled kernels and the student does not. Median replan times are 0.445 s (teacher), 0.636 s (student), 0.127 s (π0.5), and 0.389 s (Diffusion Policy). These figures do not establish a student speed advantage.

05 / Contact intervention

What changes when imagined contact is clamped?

We keep the deployed student checkpoint fixed and replace the generated contact frames with zeros throughout sampling. Thirty additional robot trials are compared with the intact student’s previously recorded outcomes on the same first ten starts per task.

Successful placements fall from 21/30 to 4/30, a decrease of 56.7 percentage points. The largest task-level change occurs on waffles.

Fixed checkpoint · matched start seeds
Intact studentZero contact
Figure 6 Placement counts out of ten per task. The intervention changes the generated representation; contact-dependent selection and closing rules can also mediate the result.

This supports deployment-time dependence on the contact-generation pathway in the evaluated system. It does not isolate ACC’s attention bias, establish contact-prediction accuracy, or compare contact distillation with action-only training. Intact trials precede intervention trials, so session-order effects remain possible.

Offline intervention results

On 124 held-out validation windows, zeroing generated contact raises the student’s endpoint error from 22.31 mm to 28.31 mm, the no-motion floor. Zeroing only the ACC input leaves it at 22.35 mm. These interventions act on different parts of the system.

06 / Contact mechanics

Placement and force are different outcomes.

We report pinch force separately from placement. The student has lower mean force than the teacher on waffles and eggs, and a similar mean on cartons. These summaries are conditional on sensor-defined placements, whose counts can differ from operator verdicts.

Pinch force at sensor-defined placements
Figure 7 Points show means; lines show reported standard deviations, not confidence intervals. Samples are shown explicitly. The one- and two-sample Diffusion Policy entries report only a mean. Missing observations are not plotted as zero.

The 24.8 N reference used for waffle/carton analysis is derived from those observations; it is not a calibrated safety limit for eggs. Two egg cracks were recorded for Diffusion Policy and none for the other models. Conditional force summaries and limited event coverage do not establish damage-free control.

07 / Discussion

What the evidence establishes.

HapticWAM provides a way to transfer structured contact generation into a policy that does not consume fingertip observations. The evaluated student has the highest observed placement counts, and clamping its contact generation changes closed-loop outcomes substantially.

The experiments are exploratory, use single training runs, and cover three tasks on one platform. They do not establish calibrated contact forecasting, an isolated benefit of ACC, or superiority to a matched action-only distillation control. The shared tactile-assisted controller remains part of the reported system.

Matched training controls, held-out contact-prediction evaluation, and interventions separated from execution rules are important next steps.

08 / Research materials

Paper, data, and models.

The report follows the manuscript revision supplied on 20 September 2026. Demonstrations and contact arrays are derived from pinned public Hugging Face archives.

Models

Datasets

Citation

@misc{sannikov2026hapticwam,
  title = {HapticWAM: Distilling Imagined Touch into a
    World--Action Model without Inference-Time Tactile Sensing},
  author = {Sannikov, Mikhail and Mikhalchuk, Ilya and
    Gubernatorov, Konstantin and Kovalev, Petr and
    Oluwatobi, Ogunwoye Faith and Tsetserukou, Dzmitry},
  year = {2026},
  eprint = {2609.23888},
  archivePrefix = {arXiv}
}

Architecture figure

Hover to magnifyPinch to zoom · drag to pan