PHASE (PHase-Aware Segmentation and REtrieval) runs in three stages. A masked tactile-proprioceptive encoder named MAT $^3$ is first trained on the prior dataset with self-supervised masked reconstruction, giving per-frame embeddings that capture force-driven interaction dynamics without any phase labels. Demonstrations are then segmented at interaction-defined boundaries and matched phase-to-phase, and an ACT policy is finally trained on the target demonstrations plus the retrieved, phase-consistent segments.

Figure 2: Overview of PHASE. The framework consists of three stages: (1) Tactile Representation Learning, where a masked tactile-proprioceptive encoder learns embeddings that capture force-driven interaction dynamics; (2) Phase-Aware Segmentation & Retrieval, where demonstrations are segmented using tactile-estimated torque signals to identify interaction-defined contact phases and variable-length segments are retrieved accordingly; and (3) Policy Training, where an ACT policy is trained on the retrieved, phase-consistent segments for robust peg-in-hole insertion.

The soft wrist is what makes segmentation possible: passive 6D compliance keeps the peg in sustained contact, so the tactile-estimated torque $\hat\tau = \sum_i r_i \times f_i$ stays smooth enough to read. A boundary is placed at the first frame after the contact peak where the coefficient of variation of $\|\hat\tau\|$ drops below $\eta$, the onset of stable engagement, splitting each insertion into a search and an insert phase. Because that moment depends on geometry and initial pose, the resulting segments are naturally variable in length, and retrieval uses FastDTW between per-frame embeddings so segments of different duration can still be aligned, restricted to matching phases.

Figure 3: Tactile-Based Phase Segmentation. The top panel shows the tactile-estimated torque magnitude (dark curve) and its signed components (light curves); the bottom panel shows the coefficient of variation (CV) of the torque magnitude. The boundary between search and insert is detected at the first frame where CV < η after the contact peak.

We evaluate on real-world peg-in-hole insertion with a UR5e, a soft wrist, and a 3×3 distributed tactile sensor, across five peg geometries: circle and square are in the prior data, while rectangle, oval, and hexagon are unseen. From 122 prior demonstrations and only 4 target demonstrations per shape, PHASE reaches 77% success, 13 percentage points above the strongest non-phase-aware baseline, and every retrieval strategy is compared under the same ACT policy architecture.

Real-world success rate (20 trials per shape, 100 per method)

The gap widens under distribution shift. Starting the robot from positions never seen in training, PHASE retains 47% success where the best baseline manages 17%.

Success rate under unseen starting positions (20 trials per shape, 100 total)