Where things stand
Researchers are building the pieces, but the evidence is lab-scale. A touch reflex on a Unitree G1 reacts within 10 milliseconds and halves contact impulse against a slower reflex, and a social-gesture system reached a 0.90 accuracy score over 150 real-robot trials. None of the papers we track report a safety certification or a long-term trial beside workers.
Safety work is moving below the main controller. Learned controllers typically update about 50 times a second, which is too slow to react to a hit, so teams are adding reflex paths that act directly on the motors.
Compliance, or giving way when pushed instead of resisting, is a growing focus, because a robot that yields is less likely to hurt someone it bumps into.
Social interaction research, such as handshakes and hugs, is still built on small groups of participants. A robot that is safe in a lab demo is not the same as one cleared to share a factory aisle.
The papers
AffectiveVLA: Safe and Intent-Aware VLA Control for Human–Humanoid Social Interaction · Nikita Kuzmin et al. (Skolkovo Institute of Science and Technology) · arXiv 2611.07712 · Nov 5, 2026 · Preprint
- What they did
- Reads what a person intends from video and responds with a handshake, fist bump or hug, with a spoken safe-stop command.
- Result
- With intent recognition added, it scored 0.90 on a combined accuracy measure (macro-F1, 0 to 1) over 150 real-robot trials: hug 1.00, handshake 0.92, fist bump 0.78. The best model without it scored 0.75. The voice stop worked in 10 of 10 trials.
- Caveat
- Training data came from 19 people, and the user study on how safe and comfortable it feels is still running.
Checked against: Paper text. Results as reported by the authors; not peer-reviewed unless noted.
ReflexArc: Bottom-Up Integration of Instantaneous Tactile Reflexes into Whole-Body Humanoid Control · Mingyu Wang et al. · arXiv 2611.04424 · Nov 3, 2026 · Preprint
- What they did
- Gives a humanoid a fast touch reflex: skin sensors trigger a pull-away motion sent straight to the motors, skipping the slower main controller.
- Result
- The reflex reaches the motors within 10 ms. Against the same reflex delayed by 20 ms, the first force spike fell 34.1% and contact impulse fell from 0.263 to 0.123 N·s, about 2.15 times lower, across three kinds of hazardous contact.
- Caveat
- Controlled lab contact tests. This is not a safety certification.
Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.
Moderate evidencePhysical robot and simulationNot named in the abstract
CompliantBot: Learning Whole-Body Compliance across Task Interfaces for Humanoid Control · Haoyang Li et al. · arXiv 2611.01748 · Nov 1, 2026 · Preprint
- What they did
- Teaches a humanoid to give way when pushed or pulled, by moving its arms, stepping or turning its body, instead of resisting.
- Result
- The authors report it yielded farther under steady pulls, with lower interaction forces, lower joint power and the highest task completion under strong pulls among the methods compared. Real-robot demonstrations are shown.
- Caveat
- The abstract gives no numbers or named baselines, and does not say which comparisons were run on hardware.
Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.