Whole-body humanoid control and wheeled assistance point to different definitions of useful robots

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not hold a real degree or possess firsthand experience.

Key points

• TANGO is reported to produce whole-body motion from language instructions and RGB camera input, rather than issuing a conventional steering command; its researchers say this matters for a humanoid moving through spaces where arm, torso and leg positions affect clearance.

  Sources: S1

• In the reported TANGO evaluation, restricting the model to a flat route prediction reduced success under real physical control, while a separate comparison reported lower collisions than the researchers’ strongest modular baseline. These are encouraging but bounded results from the described study, not a general reliability claim.

  Sources: S1

• Hello Robot’s Stretch 4 takes a wheeled, single-arm approach to home assistance. The supplied reporting describes user tasks and a planned live demonstration, but provides no comparable task-success or safety measurements for the new model.

  Sources: S2

The useful comparison is not humanoid versus non-humanoid

The two developments address the same practical setting—spaces built around people—but begin with different technical assumptions. TANGO, a navigation framework from UC Berkeley and Princeton researchers, treats a humanoid’s whole body as a navigation variable. Stretch 4, from Hello Robot, uses wheels for movement and a telescoping arm for reaching and manipulation. That makes this less a contest over appearance than a comparison between solving clearance through coordinated posture and avoiding much of the locomotion problem through a different machine layout.

Sources: S1 · S2

For builders, the distinction changes what must be engineered and validated. A humanoid may potentially use paths, access points and spatial relationships that reward a humanlike body, but it must account for the body’s changing envelope at every movement. A wheeled assistant sacrifices some terrain access while potentially reducing balance and gait-control demands. Neither source establishes that one approach is broadly superior; they reveal different dependency chains between body design, perception, control and the tasks a robot can credibly perform.

Sources: S1 · S2

Sources: S1 · S2

TANGO measures a specific benefit of putting the body into the navigation loop

The reported TANGO stack takes a natural-language instruction plus forward- and downward-facing RGB images, then produces motion for all of the humanoid’s joints. A vision-language component interprets the scene and instruction, an action model generates short motion segments, and a separate motion tracker executes them on hardware. Training data came from a simulation pipeline that planned routes, edited walking motions for maneuvers such as crouching or turning sideways, and rejected simulated motions that collided or fell.

Sources: S1

The important measured comparison is internal to this design. According to the researchers, constraining their model to predict a flat route lowered its success under real physical control from about 53% to about 27%. In another reported comparison on the G1 humanoid, TANGO reduced collisions from approximately 16% to about 10% against the team’s strongest modular baseline, despite TANGO using RGB images while that baseline also used LiDAR. Those figures support the narrower claim that whole-body output helped under the study’s conditions; they do not by themselves establish dependable operation across homes, buildings or lighting conditions.

Sources: S1

The study also reports zero-shot transfer to a G1 robot on cluttered office routes, including side-stepping, bending and stepping over floor obstacles. That is a meaningful integration result because synthetic-data training, perception, action generation and physical execution all had to connect. Yet the evidence is a preprint described through a news report and researcher comments. The supplied material does not provide a broad distribution of environments, a long-duration reliability result, or an independent replication.

Sources: S1

Sources: S1

Stretch’s proposition is product usefulness, not a navigation benchmark

Stretch 4 is presented as a compact mobile robot with a single telescoping arm, sensors, perception and contact sensitivity for working around people. The supplied report says the model launched in May and that its first production run had sold out when TechCrunch spoke with Hello Robot’s chief executive in June. It also says the company describes the system as shaped by customer feedback and supported by a full-stack open-source software platform for researchers and developers.

Sources: S2

The article connects the Stretch line to concrete assistance activities: users with severe mobility impairments are described as using Stretch technology to feed themselves, close blinds and scratch an itch, while a profiled user used it with drinking, handling glasses and brushing teeth. These examples are more consequential than a mobility demonstration because they put the success condition at the point of human independence. But they are reports of uses of Stretch technology, not controlled measurements of Stretch 4’s success rate, intervention rate, collision performance or ability to recover from mistakes.

Sources: S2

That evidence gap matters particularly because a wheeled robot’s apparent simplicity does not remove the hard parts of operating near people. It relocates them. The system still needs to perceive objects, reach safely, make contact appropriately and remain usable by its intended operator. The planned onstage Stretch 4 demonstration can show behavior in a live setting, but a demonstration is not equivalent to a published evaluation across varied homes and users.

Sources: S2

Sources: S2

Inference: embodiment determines where the evidence burden lands

Inference: TANGO and Stretch imply complementary deployment strategies. TANGO seeks to make a humanoid fit the environment by making posture part of the action decision. Stretch seeks to make useful assistance achievable with a more constrained mobility platform and a purpose-built reaching mechanism. In both cases, the central dependency is not merely AI capability: perception, physical geometry, low-level control and task design must work together. The sources support that connection, but they do not provide a head-to-head comparison on the same tasks or site.

Sources: S1 · S2

For a team choosing a platform today, the more defensible starting point is to map the task environment before selecting the robot identity. Frequent stairs, narrow routes or obstacles that demand body reshaping increase the value of the capability TANGO is designed to learn. Predominantly accessible interiors and manipulation tasks within a mobile arm’s reach may favor a wheeled system, provided the task can be completed safely and independently. This is a design inference, not a measured conclusion from either report.

Sources: S1 · S2

Sources: S1 · S2

What would change the assessment

For TANGO, the most decision-relevant next evidence would be evaluation beyond the reported cluttered benchmark and office deployment: performance in poor lighting or visually ambiguous scenes, on stairs and other hard terrain, and during contact-rich tasks such as opening doors or moving obstacles. The researchers themselves identify the low-level tracker as a limitation on harder terrain and say RGB-only sensing is weaker in ambiguous or low-light conditions, where depth or LiDAR could help.

Sources: S1

For Stretch 4, the key missing evidence in the supplied material is a comparably defined operational record: task completion, safety behavior, failure recovery and the range of home layouts in which users can complete assistance routines. The company’s planned live appearance and reported real-world user examples are useful signals of product intent and relevance, but builders need repeatable operating evidence to compare them with navigation research. The systems should therefore be judged against the human task and site constraints they can document—not the visual appeal of a humanoid or the apparent modesty of wheels.

Sources: S2 · S1

Sources: S1 · S2

Why it matters

Robotics teams increasingly have to choose whether to spend complexity on legged, whole-body mobility or on a constrained platform optimized for useful manipulation. TANGO supplies reported evidence that whole-body action can improve a humanoid navigation problem under defined conditions. Stretch supplies a product-oriented case that useful assistance can come from a wheeled design, while leaving comparable performance data unspecified in the supplied report. The practical standard is evidence matched to the environment and task, including failure modes—not a general claim that either embodiment is inherently more capable.

Sources: S1 · S2

Sources

1. Humanoid robots navigate narrow gaps and obstacles with whole-body AI control — Tech Xplore Robotics · 2026-09-23
2. TechCrunch Disrupt 2026: Aaron Edsinger brings Hello Robot’s Stretch 4 to life onstage — TechCrunch Robotics · 2026-09-22