RATE: Risk-Aware Tactile Encoding for Contact-rich Robotic Manipulation
RATE (preprint) introduces a risk-aware tactile encoding method that learns task-conditioned contact risk from tactile history and alert supervision, then injects it into a frozen ACT policy through a lightweight residual adapter. It reports 71.8% average success on eight UniVTAC simulation tasks and 73.3% average success on three real-world tasks, outperforming baselines including 3.3B-5B parameter visuomotor policies while updating only about 0.59% of model parameters. This matters for contact-rich manipulation because similar tactile readings can hide very different failure risks.
Can tactile representations explicitly encode task-conditioned interaction risk from history to improve contact-rich manipulation policies without resorting to large-scale models?
In contact-rich manipulation, similar tactile observations can arise from sensor noise, benign task-necessary variations, and early signs of undesirable contact, so downstream policies often misread task risk and either correct unnecessarily or respond too late. Existing tactile and contact-aware representations do not organize tactile features according to task-conditioned risk in the context of an evolving interaction.
Prior work used visuo-tactile contrastive or shared-latent representations (e.g., VITaL, UniVTAC), large-scale simulation pretraining, recurrent or predictive temporal tactile models, and reactive/predictive contact-aware policies. These capture contact appearance, geometry, and dynamics, but risk-related signals are used mainly for event prediction, control constraints, or action correction rather than to shape the tactile representation itself.
RATE pretrains a bilateral tactile-history encoder using a two-layer LSTM over nine tactile frames with stride five. It is trained with joint objectives: short-horizon future prediction of tactile RGB, depth, and contact-pose changes, plus future task-conditioned alert regression. The learned 512-dimensional risk-aware representation is then fused into a frozen pretrained ACT policy via a zero-initialized residual MLP applied to both left and right tactile tokens, with only the input projection and adapter trained; alert-weighted action imitation focuses learning on task-critical steps.
On eight UniVTAC simulation tasks with 100 rollouts per task, RATE reaches 71.8% macro-average success vs 47.4% for same-configuration ACT+UniVTAC retrained, 30.9% ACT vision-only, 40.5% VITaL, 41.4% π0.5, 56.1% StarVLA-α, and 16.5% GigaWorld-Policy. On the four Contact-Rich tasks, RATE averages 72.0% vs 49.0% for StarVLA-α. Ablations: base tactile 44.0%; discrete alert 56.1%; continuous severity alert 57.5%; transition-aware alert 56.0%; context-only 56.0%; full RATE 71.8%. On real robots with 20 trials per task, RATE achieves 75% USB insertion, 85% bottle-cap screwing, and 60% plug insertion success; macro-average success 73.3% vs 40.0% for π0.5. Contact-safe success rates are 93.33%, 100%, and 83.33% for RATE on the three tasks, highest in all cases. RATE has about 133M parameters, versus 3.3B-5B for large-scale baselines.
The authors state that subtle but task-relevant tactile variations may become less distinguishable when contact magnitudes vary substantially across interaction stages, and future work should explore scale-adaptive encoding. The paper is a preprint and does not report code or model release. Real-world evaluation uses only three tasks on a single xArm7 platform with 20 trials per task, and the contact-safe metric relies on maximum tactile depth as a proxy for undesirable deformation rather than direct force measurement. Simulation evaluation uses 100 demonstrations per task on eight benchmark tasks with task-specific models.
This could benefit companies building contact-rich assembly and manipulation: electronics/connector insertion, screwing and fastening, 3C manufacturing automation, and robot integrators using tactile sensors such as Daimon Robotics. Adoption could occur in 1-3 years if the approach generalizes across more tasks, grippers, and robots, and if alert signals can be generated from existing sensor streams or demonstrations; near-term use is more likely in research labs and pilot lines. Release of code/pretrained encoders and larger multi-task datasets would accelerate uptake.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.