RATE lifts UniVTAC eight-task macro-average success to 71.8% and, with 0.59% of parameters updated, raises both success and contact-safe success on three real-world tasks
Related research and updates
Synopsis
The authors propose Risk-Aware Tactile Encoding (RATE): bilateral tactile history is aggregated by a two-layer LSTM into interaction context, shaped jointly by short-horizon future contact-change prediction and continuous task-conditioned alert supervision, then added to a frozen pretrained ACT policy through a zero-initialized residual adapter that updates about 0.59% of parameters; RATE reaches 71.8% macro-average success on the eight UniVTAC simulation tasks (retrained ACT+UniVTAC: 47.4%) and, on real-world USB Insertion, Bottle Cap Screwing, and Plug Insertion, achieves success rates of 75%, 85%, and 60% with contact-safe success rates of 93.33%, 100%, and 83.33%.
Figure 1: RATE in contact-rich manipulation: corrective behavior and task performance. (a) Representative interaction sequence: when the interaction reaches critical contact, baseline methods may continue toward failure, whereas RATE enables a corrective response toward successful execution. (b) Success rates across all eight simulation tasks, comparing RATE with representative tactile and large-scale policy baselines. RATE achieves the best or jointly best performance on six of the eight tasks. (c) Success rates on three real-robot tasks, including USB insertion, bottle-cap screwing, and plug insertion, where RATE achieves the highest success rate on all three tasks.
arXivInterpretation
RATE learns a task-conditioned risk-aware tactile representation rather than only describing contact appearance or geometry. Prior tactile representation objectives were not explicitly designed to organize tactile information by task-conditioned risk, and risk-related signals were typically used for event prediction, control constraints, or policy optimization; RATE uses history-conditioned prediction to retain interaction-evolution context and continuous alert supervision to associate that context with task risk. Ablations show the shared task-specific tactile base at 44.0% macro average, continuous alert alone at 57.5%, context alone at 56.0%, and their combination in Full RATE at 71.8%.
The risk-aware representation complements rather than replaces conventional tactile features through a zero-initialized residual, giving parameter-efficient adaptation. Zero-initializing the adapter's final layer makes the fused representation initially reproduce the original tactile tokens exactly; during adaptation the conventional tactile encoder, RATE encoder, visual backbone, and ACT backbone stay frozen, with only about 0.59% of parameters updated. Using the continuous-alert configuration as the direct controlled baseline, adding context and the learned residual raises the Contact-Rich Interaction average from 49.3% to 72.0%, including Insert Hole from 32% to 71% and Insert HDMI from 18% to 42%.
On the eight UniVTAC tasks, RATE attains the highest macro-average success with about 133.2M parameters and leads across all three task categories. Against ACT+UniVTAC retrained under the same demonstrations and camera configuration (47.4%), RATE reaches 71.8% macro average; on the four Contact-Rich Interaction tasks it improves by 38, 48, 27, and 41 percentage points, and exceeds the strongest large-scale baseline StarVLA- by 15.6 percentage points. 100 closed-loop rollouts per task with 100 demonstrations per task; some baseline success rates are published results, while ACT+UniVTAC is additionally retrained with the same demonstrations and camera configuration.
On real hardware, improved task completion is accompanied by a higher proportion of contact-safe successful executions. On USB Insertion, Bottle Cap Screwing, and Plug Insertion, RATE reaches success rates of 75%, 85%, and 60% (macro average 73.3%) and contact-safe success rates of 93.33%, 100%, and 83.33% (macro average 92.22%), exceeding the strongest comparison method on each task. 50 demonstrations and 20 trials per task from the same initial conditions; contact safety is judged by maximum estimated tactile depth staying below a threshold fixed before evaluation and applied consistently across methods.
Perspective
The work targets contact-rich robotic manipulation in settings that have bilateral tactile sensing, collectable task demonstrations, and definable task-relevant alert signals; the simulation side corresponds to the eight UniVTAC tasks with a Franka Panda and two simulated GelSight Mini sensors, and the real-world side to an xArm 7 with two Daimon sensors whose native measurements are converted into GelSight-like RGB tactile images. Its value is giving a pretrained policy risk context at very small parameter cost: future observations and alert targets are used only in training, while deployment relies solely on current and historical visual, tactile, and proprioceptive observations, making it suited to lightweight upgrades of existing tactile policies.
Alert signals are automatically derived from task-relevant interaction signals in each domain and normalized, so how their definition affects cross-task transfer remains worth watching; contact safety uses maximum estimated tactile depth as a proxy for excessive deformation rather than a direct measurement of contact force; the 20 trials per real-world task limit the precision of estimates of between-task differences; and the authors note that when contact magnitudes vary substantially across interaction stages, local temporal changes modeled through feature differences may make subtle but task-relevant tactile variations less distinguishable, with scale-adaptive tactile encoding planned as future work. In addition, some baseline success rates are cited from published results and are not fully same-source with the internally retrained configuration.