Physical AI Weekly

Episode 2 · Selected 6 Nov 2026 · 4 papers

What should a robot trust?

Four research picks on safer learning, demonstration labels, navigation representations, and wearable assistance: SEVI, KiRoSe, RepresentNav, and EgoThumb.

Listen in English · 7:49

AI narration · Kokoro 82M · af_heart. Research sources are linked below.

Three takeaways

  • 01

    Separate final success from the safety cost of learning.

  • 02

    Check automatic labels against their training distribution.

  • 03

    Test visual and wearable policies beyond familiar conditions.

Selected papers & original sources

Selection date is separate from publication date. The papers below are arXiv preprints.

  1. SEVI: Safe Evidential Real-World Reinforcement Learning

    First submitted 22 Oct 2026 · 2610.31829v1

  2. KiRoSe: Using Kinematic Data for Efficient Segmentation of Robot Demonstrations

    First submitted 22 Oct 2026 · 2610.31801v1

  3. See What I Need, Dream What I Want: A Navigation Foundation Model on a Self-Constructed Representation Space

    First submitted 22 Oct 2026 · 2610.31762v1

  4. EgoThumb: A Vision-Driven Supernumerary Robotic Thumb for Autonomous Grasp and Holding Assistance

    First submitted 27 Oct 2026 · 2610.37582v1

Full transcript

0:00Welcome

Welcome to Physical AI Weekly. Episode two, selected November sixth, twenty twenty-six, asks: what should a robot trust?

Our four picks cover safer learning, motion labels, navigation representations, and wearable assistance. Three first appeared on arXiv October twenty-second; the fourth October twenty-seventh. Selection is separate from publication date.

Both language editions present the authors' evidence, which we have not independently reproduced. Listen for each task, measurement, and limitation.

0:34SEVI: a safety layer for real-world reinforcement learning

First, SEVI asks when a learning robot should reject its own proposed action. Reinforcement learning on real hardware is costly when exploration causes unsafe contact. SEVI adds an evidential safety critic: one evaluation estimates both future violation risk and uncertainty about that estimate. High risk, or insufficient evidence, triggers an offline-trained recovery policy. The task learner's objective stays unchanged.

The hardware test is USB insertion with a Franka Research three arm. Each method receives three independent training sessions, three hours each. Final policies are assessed over fifty hardware rollouts per session; unsafe events are counted throughout training.

Across those sessions, SEVI averages eighty-three percent final success and fifteen unsafe events per training run. The SQRL recovery baseline averages sixty-eight percent success and thirty-five unsafe events. These are averages over three runs, not event totals across all nine training hours.

Final success and training safety are different measurements. Two simulated tasks are evaluated separately; their results are not hardware evidence.

There is no formal guarantee of zero violations. Recovery can be inefficient, and the controller may switch repeatedly near failure boundaries. For people training policies on real robots, the useful idea is combining predicted risk with uncertainty; it is a reduction in observed failures, rather than a proof of safety.

2:11KiRoSe: segmenting demonstrations from kinematics

Second, KiRoSe asks whether robot motion can provide useful labels for demonstrations. Manually marking the start and end of subtasks is tedious. Processing long, synchronized camera recordings with a language model can also be expensive.

KiRoSe learns a compact representation of the robot's kinematics. An encoder is trained to recognize subtask transitions. Its features become discrete Sparse Motion Tokens, which a language model reads to produce segment boundaries and timed action descriptions. This is trained annotation, rather than an annotation system requiring no supervision.

Training uses five thousand four hundred seventy-eight episodes from the YAM bimanual robot. On two hundred held-out episodes from that same robot, boundary F one reaches zero point six three when a predicted boundary may differ from the annotation by up to one second. Alternative kinematic tokenizers using the same reader reach roughly zero point four one to zero point four two.

The authors also evaluate another robot dataset and human-hand trajectories. Their reversal test suggests that motion features can track decreasing progress when an action is played backwards. That does not mean KiRoSe wins every comparison: competing progress models perform better on forward prediction, and accuracy is lower on human hands.

For teams organizing demonstrations, the compact motion representation is worth studying. The authors still recommend spot-checking labels, especially far from the training distribution. Automatic annotations should not replace human review of safety-critical failure and recovery data.

3:59RepresentNav: a representation the task builds itself

Third, RepresentNav asks what a navigation policy should see before deciding where to move. A pretrained visual encoder preserves information useful for its original training objective. That may differ from what navigation needs.

RepresentNav combines features from multiple depths of three frozen vision foundation models. A small, three-layer transformer learns a representation together with the policy. Before acting, the system predicts a future representation in that same learned space. This is a prediction of visual features, rather than a guarantee that a simulated future will occur.

On VLNVerse with fine-grained instructions, the authors report eighty-three point eight percent success, twenty point zero five percentage points above the previous best method reported in that comparison.

A separate point-goal analysis tests a more specific question: where should the future prediction live? Predicting in the policy's learned space reduces collision rate from eighteen point zero five to thirteen point four five percent. Predicting the backbone's pooled features instead raises it to twenty-two percent. These collision rates come from that separate protocol, not the VLNVerse success-rate comparison.

The construction has been validated on navigation. Transfer to manipulation, dexterous hands, or humanoid whole-body control remains untested. For people designing visual policies, the practical question is whether the representation is optimized for the task that will actually use it.

5:36EgoThumb: a wearable robotic thumb

Fourth, EgoThumb brings the question to a device worn beside a human hand. In a hold-and-manipulate task, the extra robotic thumb holds an object while the person's fingers work on it. Tendon-driven actuation sits on the forearm. A palm-facing camera observes the hand and object, and a visual imitation policy maps those images to joint positions for grasping, holding, and release.

The authors describe an open-source system. Their evaluation includes three participants, four tasks, and ten trials per condition. With objects inside the initial and target areas represented in training, average success across tasks ranges from eighty-seven point five to ninety-two point five percent across the three participants.

Moving beyond those areas exposes a substantial limit. On Transfer Ball, with expanded initial and target areas, average success falls to seventeen percent. This is one task under that expanded-area condition, not the success rate of the entire device.

Longer tasks and larger spatial variation are harder. Fast hand movements can blur the camera image, high-speed manipulation remains unevaluated, and varied hand-approach angles have not been established. The protruding camera may also affect comfort.

For wearable-robotics teams, the useful contribution is the coordination between the camera, learned control, and a simple assistive mechanism. Three participants and binary task outcomes still leave substantial work before broad claims about everyday assistance.

7:14Takeaways

Trust is the common thread: risk and evidence, motion labels, task-specific vision, and hand-level observations.

Our interpretation stays within the tests. Strong averages can hide hard conditions; final success can hide learning costs.

Start with SEVI for real-robot learning, KiRoSe for annotation, RepresentNav for navigation, and EgoThumb for wearable assistance. Each guide links to the pinned paper.

This briefing uses synthetic narration. Thank you for listening; keep the limitations in view.

Feedback on pronunciation, translation, or content: Leave feedback on GitHub