PHASE retrieves tactile insertion experience one contact phase at a time
OMRON SINIC X and University of Tokyo researchers segment insertion demonstrations by tactile-proprioceptive contact phase, then retrieve matching prior segments for few-shot policy learning.

Researchers from OMRON SINIC X and the University of Tokyo released PHASE on September 25, 2026. Accepted at IROS 2026, PHASE uses a compliant wrist and tactile-proprioceptive signals to split insertion demonstrations into contact phases, retrieve relevant portions of prior experience and adapt a new policy using four target demonstrations alongside selected experience from a 122-demonstration prior. Official project page.
Explore tactile sensing for the broader comparison.
Key takeaways
- With four target demonstrations per peg shape, PHASE reaches 77% success in the normal-start evaluation, 13 percentage points above the strongest 64% comparison.
- Under unseen starting positions, PHASE reaches 47%, compared with 17% for the strongest baseline—a 30-point gap under the reported setup.
- The evaluation uses one UR5e, one compliant wrist, a 3-by-3 tactile array and 20 trials per shape. Code and data were not publicly linked.
What changed
Demonstration retrieval usually compares entire trajectories. PHASE argues that contact-rich insertion is better compared in stages. A masked tactile-proprioceptive encoder called MAT3 represents contact, while a soft wrist makes the phases physically observable. Offline segmentation places a boundary after the full demonstration’s torque peak when a subsequent stability condition is met, with a fallback to the interval end. This is not a causal runtime detector that knows the future peak.
Fast Dynamic Time Warping then compares like phases between four target demonstrations and a pool of 122 prior demonstrations. The retrieved segments train an Action Chunking with Transformers policy. This avoids forcing an entire prior trajectory to match when only its search or insertion segment resembles the target. Method and paper links.
Results under the reported conditions
The physical setup uses a UR5e arm, a soft compliant wrist, a 3-by-3 PapillArray, wrist force-torque sensing and HTC VIVE gripper-pose tracking. Circle and square pegs supply the 122 prior demonstrations. Rectangle, oval and hexagon are absent from that prior but still have four target demonstrations each: this is few-shot adaptation, not zero-shot transfer. Target-only and retrieval policies use the target examples; BC-Prior is trained only on the prior. Each method has 20 trials per shape, 100 in total.
Under the normal-start protocol, the target-only policy reaches 50%. Training on all prior data without retrieval reaches 64%; single state-action-pair retrieval reaches 56%; whole-trajectory retrieval reaches 64%; fixed-window retrieval reaches 55%; PHASE reaches 77%. Success means full insertion within 60 seconds. The relevant aggregate gain is 13 percentage points over the strongest 64% baselines, not a 13% relative improvement or superiority on every peg shape.
When start positions are perturbed by up to plus or minus 2 cm along each horizontal axis outside the demonstrated starting poses, target-only falls to 0%, the no-retrieval prior baseline reaches 17%, single retrieval 2%, full retrieval 14%, fixed-window retrieval 16% and PHASE 47%. The 30-point margin over 17% suggests that phase-level matching helps most when a policy must first recover contact before insertion.
What this means for robotics
RoboSkin analysis: PHASE treats compliance as part of perception. The soft wrist changes the measured response to contact, making the phases more observable in the authors’ setup; that does not establish fixture protection or force safety across assemblies. That coupling between mechanism and learning is relevant to robot manipulation systems that cannot rely on vision after the peg enters a fixture.
The work also offers a practical data strategy. Instead of discarding old demonstrations because a new peg is different, teams can reuse the segments whose contact dynamics match. But the reuse benefit depends on recognizable phases. Tasks with repeated impacts, slip or more than two stable contact stages may require a learned or multi-boundary segmenter.
Limitations and availability
PHASE is an IROS 2026 conference paper with an arXiv v1 record, and RoboSkin.ai has not reproduced it. Results come from a single arm, wrist and 3-by-3 tactile configuration. Twenty trials per shape limit precision, and the five peg geometries do not establish transfer to threaded, flexible or high-force assembly.
The official project page provides paper, poster and video links. It does not link public code, trained policies or a demonstration dataset, and no software or data license was verified. The earlier OMRON SINIC X announcement remains a source route, but returned an access restriction on the November 4 recheck. The accessible author paper and project identify IROS 2026; none of these constitutes independent experimental validation.
Distinguish offline segmentation from runtime observation
The retrieval representation and the deployed policy do different work. PHASE’s offline bidirectional encoder can use expert-action tokens from complete demonstrations. The runtime policy does not receive those future expert actions. Phase matching prepares training examples; it is not a claim that the robot retrieves and chains prior segments online. Methods and evaluation.
| Training component | What to preserve when adapting it |
|---|---|
| Torque-based segmentation | Signal definition, peak/stability rule and fallback cases; heuristic termination is not verified boundary-label accuracy |
| Retrieved prior segments | Segment identity, phase alignment and masks; train each as an independent trajectory rather than crossing artificial boundaries |
| Complete target demonstrations | Their real search-to-insert transition and target-specific evidence, even for shapes absent from the prior |
| Runtime observation | Tactile, wrist and pose channels, their clocks and frames; do not assume tactile-only deployment |
The best aggregate score also hides object differences: PHASE records 18/20 on circles versus BC-Prior’s 19/20, and 14/20 on hexagons versus single-pair retrieval’s 15/20. Its overall 77/100 is useful evidence, but not a guarantee for each geometry. The short demonstration durations are separate from the 60-second evaluation limit.
For a new fixture, record protective stops, timeout, drift and recovery as well as insertion. The paper identifies those failures and does not supply a general uncertainty estimator or large-offset recovery policy. Use the manipulation decision map to distinguish a better training-data selection method from measured execution reliability.
Sources and related resources
- PHASE arXiv v1, submitted September 25, 2026
- Official PHASE project page, paper, poster and video
- OMRON SINIC X IROS 2026 announcement
- RoboSkin tactile sensor benchmark guide
- RoboSkin tactile manipulation hub
*Source review updated November 4, 2026. Original publication dates are retained. Results belong to the cited authors; RoboSkin has not run these systems or downloaded their experimental payloads.*