Robotics Briefing /
Humanoid navigation gets a whole-body test.
A clear path on a floor plan can still leave a robot's shoulders stuck. Today's research examines how robots fit through clutter, turn camera views into useful motion, and capture the touch signals behind skilled hand movements. New industry announcements show where those ideas could meet practical deployment.
Your quick takeaways
- TANGO coordinates a humanoid's whole body during navigation. It completed 10 of 15 physical trials in clutter, with fewer collisions than its comparison system, but contact with obstacles remained.
- 3DWay uses two camera views to improve manipulation, while TacClip records human touch without covering the fingertips. Both address physical information that images alone can miss.
- Antioch raised $32 million for simulation and evaluation. FANUC and Palladyne AI announced an industrial collaboration, and Dynamic Creatures is bringing mobile character robots into hospitality development.
Papers
TANGO: a humanoid plans around its whole body
Anqi Li and colleagues • Submitted September 8 • Preprint; simulation and physical-robot tests
Walking through a narrow gap requires more than choosing a route. TANGO uses camera images, language instructions, and joint positions to predict coordinated whole-body motion. The system trains on simulated traversal examples, then sends its predictions to a separate controller that tracks the movement on a Unitree G1.
In physical cluttered-navigation tests, TANGO completed 10 of 15 trials, compared with 6 of 15 for InternVLA-N1 with Unitree control. Mean collisions fell from 1.93 to 0.73 per trial. Those counts make the result more informative than the demonstration videos alone: the robot improved, but still hit obstacles.
The experiment supports learning body clearance as part of navigation. Limits include three scenes per setting, instructions that explicitly name maneuvers such as bending or stepping over, and a networked GPU server for the large model. The authors promise code, data, and model checkpoints; the project page currently offers demonstrations without those release links.
3DWay: two views turn image points into robot motion
Ziqin Huang and colleagues • Submitted September 8 • Authors report ECCV 2026 acceptance; physical-robot tests
A point in a camera image leaves an important question unanswered: how far away is it? 3DWay predicts matching waypoints in two views, then uses camera geometry to recover positions in three dimensions. These intermediate motion targets can guide an existing vision-language-action model, which connects images and instructions to robot actions.
On an AgileX PIPER arm, adding 3DWay to a fine-tuned π0 policy raised average success across five basic tasks from 21.7% to 65.8%. Training used 20 demonstrations per basic task; evaluation used 24 trials per task. Tests with unfamiliar objects and changed language instructions also improved, though success remained far from universal.
The useful lesson is that explicit geometry can help a robot use limited task data. The method requires calibrated cameras and mainly represents position, leaving difficult rotations and dynamic adaptation unresolved. The public repository contains training scripts and a waypoint-processing package. The reported gains still need independent replication.
TacClip records touch while leaving the fingerpad bare
Yuqian Ye and colleagues, Stanford and Ant Group • Submitted September 8 • Preprint; wearable-sensor experiments
Human demonstrations can show a robot where a hand moved while missing the forces that made the task work. TacClip clips over a fingernail and measures tiny deformations through an optical fiber. The fingerpad stays exposed, preserving the person's direct contact with fabric, tape, or other surfaces.
After calibration, the prototype estimated contact-force magnitude with typical errors below 0.5 newtons across the tested 0–8-newton range. Physical experiments captured cloth texture and transparent tape edges, plus an underwater force response after temperature compensation. These are sensing results; the study does not show improved robot learning.
Richer demonstrations could eventually explain how people adjust force during delicate tasks. TacClip still needs broader user testing and calibration for fit, finger properties, and temperature. It estimates overall load, without separating directional forces. The authors plan to release CAD files and software after review.
News
Antioch raises $32 million for continuous robot testing
Announced September 8 • Series A; company and customer claims
Antioch announced $32 million in new funding, bringing its disclosed total to $40.5 million. Its platform combines explicit physics and geometry with learned behavior to test complete autonomy systems across many simulated conditions. NVIDIA simulation tools and Nebius infrastructure support the approach; an Amazon executive describes agreement with physical tests held out from calibration.
For readers following robot reliability, the promise is quicker discovery of failures before a software change reaches hardware. The announcement provides customer testimony, without a public benchmark of prediction accuracy or development-time savings. How well simulated failures match physical ones remains the key measure.
FANUC and Palladyne AI target more adaptable industrial robots
Announced September 8 • Planned collaboration; company claims
FANUC America and Palladyne AI will work on combining FANUC robots with the Palladyne IQ software platform. Their plans cover adaptive motion, human-assisted learning, simulation, and validation of manufacturing and logistics tasks. They also aim to standardize deployment workflows for system integrators.
The potential benefit is less custom engineering when products or tasks change. This release describes planned development and validation; it supplies no measured deployment savings or customer performance results. It is a separate announcement from FANUC's earlier IMTS preview.
Dynamic Creatures brings mobile characters into hospitality development
Announced September 8 • Company launch and partnership
Dynamic Creatures emerged from stealth as Boston Dynamics' official entertainment and hospitality partner. Boston Dynamics says the company is working with a large theme park and retailer on mobile characters that respond to guests in real time. Former leaders from Boston Dynamics, Google, and the Robotics and AI Institute founded the company.
These environments could test how movement and social interaction work together around people. The announcement names no customer, rollout date, or measured operating performance. Public operating data would clarify how well these interactions hold up during a full day.
Highlights
Orbbec puts human data collection on display in Tokyo
Announced September 8 • Logis-Tech Tokyo showcase; company validation claims
Orbbec is showing its full robot-free data-collection lineup in Japan for the first time, including first-person cameras, handheld manipulation devices, wrist cameras, and synchronization hardware. The company reports synchronization error below one millisecond under its validation conditions.
The connection to TacClip is useful: recording human activity can supply robot training data, but timing, viewpoint, and contact information all affect its value. This is a regional showcase of the lineup, with no new robot-policy performance result in the announcement.
Read this first
Read TANGO's Table 3 alongside the project demonstrations. The table separates task completion from collisions and shows the small physical trial counts. That combination gives a clearer picture of progress toward navigating occupied spaces than a successful video alone.
Read the paper