Tactile In-Context Learning for Robot Agents: Tactile-ETA vs. GPT-6 Astra

In-context learning conditions a model through examples supplied at inference; it does not mean updating model weights. The useful comparison is between two ways to package experience and connect a general model to a robot. Neither study establishes that an unconfigured chat model can control arbitrary hardware.

Tactile-ETA’s main configuration identifies gpt-6-astra with low reasoning effort. TacTalent’s method configuration identifies gpt-6-astra with xhigh effort. Shared model naming therefore does not imply matched inference settings, camera inputs, prompts, action budgets or execution code.

For the distinction between training, demonstrations and execution feedback, see the robot learning guide.

Table 1 of Tactile-ETA uses 100 rollouts per task and condition, with matched initial conditions. The comparison below keeps the same two expert demonstrations, current sensory inputs and execution interface. It changes whether tactile observations are included in those demonstrations. Both columns receive current vision and touch.

The mean gain is 8.25 percentage points, not a uniform gain on every task: Lift Can changes from 3/100 to 2/100. The native Lift Can success predicate is a pose condition; it should not be reinterpreted as a separate test of stable suspended lifting.

The separate Astra study evaluates eight UniVTAC tasks and eleven OpenViTac tasks. Each input setting has 210 sampled evaluation episodes: 80 on UniVTAC and 130 on OpenViTac. Most tasks have ten episodes; Insert Block v2 has thirty, covering three shapes. The table reports task-equal mean success, not a pooled 210-trial success rate.

The vision-only condition omits tactile images from both demonstrations and online observations. V+T supplies them in both places. This comparison measures their combined contribution within this setup; it cannot identify how much comes from historical demonstrations versus current touch.

Tactile-ETA’s hardware study uses a reBot Arm B601-RS, a vitai GF225 tactile gripper, an Orbbec wrist camera and a RealSense D435i third-person camera. Both conditions receive current vision, touch and robot state. The intervention is one whole multimodal human demonstration per task versus no demonstration, with fixed model parameters.

The ten-trial headline means five Wipe Board trials plus five Pick Up Chips trials in each condition: twenty trials across both conditions. It does not mean ten trials per task. It also does not isolate tactile demonstration content from the visual and motion information added alongside it.

Tactile-ETA Table 3 changes only the tactile representation inside the same two Lift Bottle demonstrations. Action segments, visual observations, motion records and current sensing remain fixed. Every condition has 100 rollouts. “No demo tactile” still has current tactile sensing.

The selected representation always keeps the action endpoints. When a change threshold is reached, it also retains the single adjacent-frame pair around the largest normalized bilateral change. Both fingertips use the same selected times; after deduplication, each fingertip has at most four observations per segment. This is more specific than selecting arbitrary interesting frames.

This is a proposed RoboSkin evaluation protocol, not an experiment already run. Start with the published Lift Bottle task to check the reported ablation, then repeat on the other tasks before making a broader claim. First compare frame selection within one fixed controller setup; compare two controller stacks only in a separately documented study.

Reuse the same two source demonstration episodes, action segmentation, visual images, robot-state records and action records. Keep the visual image bytes, resolution, camera views and ordering identical across conditions. Keep the online observation policy and tactile packaging fixed. Online trajectories can diverge after different actions, so actual later camera frames cannot be guaranteed identical in closed-loop trials.

Use the tactile experiment worksheet to record planned conditions and evidence. Add fields for task, scene seed, repetition, condition, demonstration IDs, selected timestamps, model/settings, prompt size, latency, outcome and failure reason; leave outcome fields empty until a run occurs.

Use this page to select a context representation and identify the controls missing from a fair experiment. Cite the original paper and its exact table for a numerical result; cite this page for the cross-study interface comparison and proposed evaluation design. Both research sources are v1 preprints, so later revisions may change their claims.

The linked repositories were inspected at the revisions below on October 22, 2026. Source inspection establishes what the released materials describe; it does not establish a successful installation, controller execution, measured latency on your system or physical robot reproduction. The TacTalent results are the research authors’ evaluation, not an OpenAI robotics certification or benchmark guarantee.

Neither a simulator’s controller frequency nor a stream’s sampling rate is the model’s decision frequency. Tactile-ETA makes decisions between motion segments; TacTalent executes generated action sequences through its adapter. Hardware integration still needs an appropriate controller, coordinate conventions, calibration and independently verified execution behavior.