Latest / Embodied AI 101 / TANGO: Stop Treating a Humanoid Like a Disk on the Floor
Transcript
- TANGO: Stop Treating a Humanoid Like a Disk on the Floor. TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model • Anqi Li, Yuxin Chen, Zhaobo Li, Zhuo Cao, Junli Ren, Masayoshi Tomizuka, Dhruv Shah • University of California, Berkeley; Peking University; Tsinghua University; The University of Hong Kong; Princeton University • Conference on Robot Learning 2026 / arXiv • 2026. Publication note, September 9, 2026: The supplied link resolves to the official TANGO project page. The paper is not from Stanford, and Chris Paxton is not an author. TANGO was posted to arXiv on September 8, 2026 and accepted to CoRL 2026. Because the main conference is scheduled for November 9–11, 2026, it is more precise to say that the paper has been accepted for presentation rather than already presented. The abstraction failure TANGO is targeting. Most vision-language navigation systems treat the robot as some variation of a disk moving across a floor plan. The model predicts a waypoint, a heading, or a discrete command such as move forward and turn left. Another system then translates that abstract route into executable motion. This decomposition has worked reasonably well for wheeled robots, simulated agents and, with additional machinery, quadrupeds. For a humanoid moving through human clutter, however, the abstraction leaks almost immediately. Consider three passages with the same valid centerline. One contains a waist-high bar, one is narrower than the robot’s frontal shoulder width, and one has a low shelf extending overhead. At the level of a two-dimensional path, all three can look identical. At the level of humanoid configuration space, they demand fundamentally different solutions: lift the feet and lengthen the stride, rotate the torso and crab-walk sideways, or lower the pelvis and reorganize the upper body. The route and the body configuration cannot be selected independently because the robot’s occupied volume changes continuously with its posture. This is TANGO’s central argument: for humanoids, navigation is already a whole-body control problem. The navigation policy should not merely decide where the pelvis travels and hope that a downstream controller finds a collision-free realization. It should produce a body trajectory whose arms, torso, legs and orientation are coordinated with the geometric constraints of the route. The resulting system, short for Traversability-Aware Vision-Language Navigation, maps language, egocentric vision and proprioception to chunks of whole-body reference motion. The authors position it as the first vision-language navigation framework to directly predict 29-degree-of-freedom humanoid joint-space actions for cluttered traversal. The deeper contribution is not simply adding more output dimensions to a VLA. It is changing the semantic meaning of a navigation action from “go that way” to “move the entire articulated body through this part of space.”. What the policy actually controls. At each decision point, TANGO receives a natural-language instruction, a temporal history from front-facing and downward-facing RGB cameras, and the robot’s current joint configuration. The two camera views are vertically stacked before entering the vision-language model. The downward view is especially important because objects that constrain foot placement can be difficult to observe from a conventional head-height navigation camera. The policy predicts a horizon of reference actions rather than one instantaneous command. Each step contains desired values for the Unitree G1’s 29 articulated joints and a six-dimensional representation of base rotation. During training, the target is stabilized further by expressing yaw relative to the start of the chunk and adding auxiliary planar displacement and heading-change targets. Those auxiliary targets make translation and turning easier to identify in a high-dimensional joint trajectory, where locomotor progress might otherwise be a weak learning signal. “Direct joint-space prediction” needs one important qualification. TANGO does not emit motor torques, and the 7-billion-parameter VLA is not closing the balance loop. It predicts executable whole-body references, which a pretrained tracking controller converts into high-frequency commands. TANGO is therefore end-to-end across language, perception and whole-body motion intent, but it remains deliberately modular at the dynamics and stabilization layer. That distinction is a strength rather than an embarrassment. Asking a large visual model running every half second to stabilize contacts at hundreds of hertz would be an unnecessarily difficult systems design. TANGO instead makes a clean division: the VLA chooses a physically meaningful future motion, while a controller with the right inductive biases handles balance, tracking and disturbances. PET: Manufacturing traversability data. The hardest part of this project may not be the VLA architecture at all. It is obtaining demonstrations that jointly contain long-horizon language navigation, realistic egocentric observations, collision-free whole-body adaptation and dynamically executable humanoid motion. Recording enough of this data on physical robots would be slow and risky. Human motion capture followed by retargeting introduces a different problem: the human demonstrator does not share the robot’s kinematic limits, collision geometry or controller errors. TANGO addresses this with an offline data engine called Plan–Edit–Track, or PET. Rather than collecting human demonstrations, PET composes classical planning, learned motion generation, geometry-aware kinematic editing and reinforcement-learning-based tracking into a synthetic expert. The source environments come from VLNVerse and SAGE-3D. The authors begin with 263 VLNVerse scenes and 1,000 SAGE-3D scenes, then use Gemini 2.5 Flash to reject scenes with problems such as incomplete geometry or unsuitable floors. This leaves 205 VLNVerse scenes and 373 SAGE-3D scenes. They augment those 578 source environments with trajectory-conditioned obstacles designed to trigger three motion families: stepping over low objects, moving sideways through narrow openings, and squatting beneath overhead geometry. The inserted assets are semantically matched to the scene—cushions and stools for low obstacles, furniture and plants for narrow passages, and lamps or curtains for overhead constraints. The phrase “trajectory-conditioned” matters. Obstacles are not sprinkled randomly through the room. Candidate objects are placed along otherwise valid navigation routes, away from the start and goal, and aligned with the direction of travel. Placement is checked against the occupancy map and other trajectories so that an augmentation creates a meaningful whole-body challenge without arbitrarily corrupting the entire scene. Across the generated environments, roughly 44 percent of targeted obstacles induce strides, 15 percent induce sideways motion and 41 percent induce squatting. The corresponding language instructions are also updated to mention the added obstacle and desired behavior. Plan. Given a start and goal, PET first runs A-star over a two-dimensional floor grid. Its path cost includes a soft penalty based on distance from occupied cells, encouraging routes with greater clearance without converting the robot into an unrealistically large fixed disk. The result is still a planar path, but a more conservative one. PET then searches that path for narrow segments by measuring lateral free space. If a passage is too narrow for frontal traversal, a rule-based heading editor rotates the desired body orientation by 90 degrees relative to the travel direction. The sign is chosen so that the leading shoulder remains on the more open side. Smooth transitions into and out of this heading produce a crab-walking reference in which body orientation and translational direction are decoupled. The path and edited heading are converted into forward velocity, lateral velocity and yaw-rate commands. These drive SONIC, a general-purpose humanoid motion system, to produce a natural 29-DoF walking trajectory. A proportional-derivative correction keeps the realized base near the planned route. At this point, PET has a dynamically plausible gait that follows the floor plan and can turn sideways, but it still does not understand overhead bars or objects that must be stepped over. Edit. The edit stage transforms that nominal walk into an obstacle-negotiating whole-body motion. PET constructs a three-dimensional signed distance field around the path and restricts editing to a corridor approximately half a meter wide. Clutter that the global planner already routed around is therefore ignored; only geometry the humanoid is expected to negotiate affects the body trajectory. For the torso and upper limbs, PET borrows the humanoid potential-field idea. Repulsive directions are sampled for links such as the shoulders, elbows, wrists and torso, then converted into bounded soft displacement targets inside a whole-body inverse-kinematics solver. A forward-progress component prevents the robot from merely retreating from every obstacle. Because the forces are soft rather than hard collision constraints, the IK solver can trade off clearance against posture tracking, contacts and center-of-mass stability. Overhead obstacles receive special treatment. Look-ahead probes detect low geometry before the head is already underneath it, while a virtual head barrier creates a downward correction as clearance shrinks. That correction is coupled to waist pitch, hip flexion and a lower center-of-mass target. The implementation can produce a waist bend, a vertical body drop, or a coordinated full squat involving the pelvis, knees and ankles. Reference-tracking weights on the upper body are relaxed during the maneuver so that the upright walking prior does not fight the crouch. Low obstacles require explicit footstep editing. At each foot lift, PET scans ahead, classifies obstacles that are low enough to step over, and retargets the touchdown beyond the obstacle. It raises the swing trajectory using an asymmetric arc that reaches clearance early, which is particularly useful for a trailing leg crossing a bar soon after lift-off. Knee-flexion bias, pelvis redirection and center-of-mass correction turn the result into a forward crossing rather than an in-place lunge. Two IK passes then reconcile the new foot contacts, support polygon, center of mass and upper-body avoidance forces. Track. Kinematic feasibility is not dynamic feasibility. An edited squat can be too fast, a swing leg can demand impossible acceleration, and a motion that clears the mesh in IK can still fall under physics. PET therefore replays each edited trajectory through SONIC in MuJoCo. Executions that fall or collide are discarded. Interestingly, the successfully tracked motion is not used directly as the action label. The authors retain the edited reference trajectory as supervision, arguing that tracker output can degrade the human-like structure of the motion. SONIC acts as an executability filter: if the controller can realize the reference safely, the cleaner reference is admitted to the dataset. The scene is then re-rendered from the verified motion, and its language instruction is regenerated. The final dataset contains 64,633 trajectories. PET motion generation consumes 86 RTX PRO 6000 GPU-hours and rendering consumes 125, for 211 GPU-hours before VLA training. This is not negligible, but it is a compelling form of privileged-information distillation: use maps, exact collision geometry, signed distance fields and physics offline, then train a policy that needs only RGB and proprioception at deployment. Three systems, three time scales. TANGO organizes inference as a three-level stack. System 2 interprets the instruction and visual history. System 1 generates future whole-body motion. System 0 tracks that motion against the robot’s physical state. This resembles the slow-reasoning and fast-execution split seen in recent VLAs, but it adds a still-faster stabilization layer. System 2 is based on Qwen2.5-VL-7B and warm-started from InternVLA-N1, giving it a navigation-specific initialization rather than beginning with a generic visual-language model. Qwen2.5-VL contributes native image and long-video processing, while InternVLA-N1 contributes learned navigation priors from its dual-system training. Long-horizon navigation creates an awkward memory problem. Retaining every high-resolution frame eventually overwhelms the context, but keeping only the latest view erases route history. TANGO uses Budget-Aware Token Sampling: recent observations are more likely to be retained and are represented with finer spatial grids, while older observations receive a lower but nonzero sampling probability and coarser tokenization. The resulting visual tokens and language instruction produce a context representation for the action generator. System 1 is a multimodal diffusion transformer trained through flow matching. In plain language, it learns how to transform a noisy candidate action sequence into a coherent trajectory conditioned on visual-language context and current proprioception. Generating a chunk rather than an isolated action provides temporal structure, but it also creates a handoff problem: if successive chunks disagree at their boundary, the robot can pause, jerk or abruptly change strategy. TANGO therefore uses training-time real-time chunking. During training, a random prefix of the target chunk is treated as already committed, and the model learns to complete the remaining horizon. At deployment, inference for the next chunk overlaps with execution of the current one, while the committed prefix anchors continuity. This moves the chunk-transition problem into the training distribution instead of relying only on post-hoc smoothing or expensive inference-time guidance. The vision-language backbone and action expert are jointly tuned. A text-decoding branch is trained on VideoQA examples to preserve general visual-language knowledge, while the flow-matching loss trains action generation. The authors also rebalance rare but important motions—large turns, sideways walking, squats and strides—and increase loss weights on the joint dimensions most responsible for each behavior. Full-parameter training runs for about seven hours across 128 A100 GPUs, or 896 A100 GPU-hours. System 0 is the SONIC tracker. SONIC was itself trained at much larger motion scale, using over 100 million retargeted motion frames and a controller of up to 42 million parameters. This prior is what lets TANGO treat whole-body trajectories as an action interface without relearning balance for every navigation task. TANGO supplies intent; SONIC supplies the high-bandwidth physical competence needed to realize it. The deployment stack. In simulation, TANGO splits visual fidelity from control fidelity. MuJoCo runs the low-level tracker and robot dynamics. A synchronized digital twin is teleported inside Isaac Sim to render photorealistic camera observations from the pose produced by MuJoCo. This avoids choosing between a fast control simulator and a visually rich navigation environment. The real system uses front and downward RealSense cameras on a Unitree G1. Images and proprioception are sent with roughly 20 milliseconds of communication latency to an RTX PRO 6000 server, where the VLA runs once every 0.5 seconds. Each prediction contains 15 actions at 30 hertz. Those references are returned to the robot, resampled to 50 hertz and tracked by SONIC on a Jetson Orin NX, with the final control loop running at approximately 200 hertz. This architecture is better described as local server–robot inference than self-contained onboard autonomy. It demonstrates real-time deployability, but the 7B perception-and-action stack still relies on workstation-class compute connected over a network. Reading the results carefully. On the original VLNVerse benchmark, TANGO is trained on 3,963 trajectories and evaluated on 423 seen and 825 unseen trajectories. The paper reports the highest success rate and lowest navigation error on both splits, with path-efficiency and oracle-success results comparable to the strongest trajectory-prediction baseline. That is impressive because TANGO executes through physics while the compared VLN methods are evaluated by teleporting an abstract agent between predicted poses. It is also not a perfectly controlled comparison: the methods differ both in policy quality and in execution model. The more revealing test is augmented VLNVerse-unseen, where routes require whole-body obstacle negotiation. TANGO reaches a 43.75 percent success rate, 31.83 SPL and 9.90 percent episode collision rate. The strongest modular baseline—fine-tuned InternVLA-N1 executed through HumanoidPF—reaches 41.88 percent success, 29.49 SPL and 15.81 percent collision rate. TANGO therefore produces a modest improvement in completion but a substantial reduction in collision incidence, roughly 37 percent relative to that baseline. It also uses RGB at inference, while the HumanoidPF executor receives LiDAR-derived geometric information. The real-world evaluation covers short routes of approximately 10 meters, longer routes of approximately 30 meters, and cluttered trials containing one major three-dimensional obstacle. Each category includes three scenes and five trials per scene, giving 15 trials per method and setting. The paper reports that TANGO completes more trials with fewer collisions than a fine-tuned InternVLA-N1 plus Unitree-controller baseline. Demonstrations show the robot navigating long corridors, turning sideways through narrow passages, ducking beneath obstacles and stepping over low bars without real-world navigation training. The ablations expose what is carrying the system. Removing real-time chunking collapses success from 43.75 to 10.94 percent. Removing motion editing reduces success to 36.25 percent and raises collision rate from 9.90 to 20.60 percent. Replacing SONIC with the independent ScaleBFM tracker gives 40.94 percent success, 29.65 SPL and 9.10 percent collision rate. The first result says smooth action continuation is not an implementation detail; it is essential to task completion. The second validates the synthetic obstacle-editing pipeline. The third suggests that TANGO’s whole-body reference interface is not completely tied to one controller. The planar action ablation provides the broader lesson. Policies that look competent when their predicted poses are teleported into the environment degrade substantially once an actual humanoid controller must realize those poses. The gap between a valid base path and an executable body motion is precisely the gap TANGO is designed to model. The closest relatives. TANGO is best understood as a synthesis of several research threads rather than an isolated architecture. From modern vision-language navigation, it inherits long-horizon visual memory and a slow semantic planner. InternVLA-N1, its initialization, separates pixel-goal reasoning from a high-frequency navigation policy and was trained using a navigation corpus containing more than 50 million egocentric images. TANGO keeps the broad dual-system philosophy but replaces the low-dimensional navigation interface with whole-body reference motion. From VLNVerse, it inherits the push toward physically embodied evaluation rather than “ghost-agent” teleportation. TANGO goes further by making the embodiment part of the learned action representation rather than merely asking a downstream controller to execute a conventional trajectory. From HumanoidPF, TANGO inherits a geometric representation of how different body links should move away from obstacles. The key inversion is that HumanoidPF uses this representation online inside an RL traversal policy, whereas TANGO employs potential fields primarily as an offline motion editor and teacher. Exact geometry generates demonstrations; an RGB-conditioned VLA then attempts to internalize the corresponding behavior. Finally, TANGO belongs to the growing family of humanoid VLAs represented by systems such as WholeBodyVLA, LeVERB and PhysiFlow. Many of those systems focus on loco-manipulation and use latent motion interfaces or high-level locomotion commands. TANGO is navigation-first and predicts explicit whole-body references, while delegating physical realization to a general tracker. What the paper proves—and what remains open. The strongest claim supported by TANGO is that whole-body action representation improves physically executed humanoid navigation. It does not prove that a large VLM can replace motion planning or low-level control. In fact, TANGO’s success depends on an excellent planner-and-editor during data generation and an excellent tracker during deployment. Its novelty lies in distilling the former into a visual policy and connecting that policy cleanly to the latter. The language-conditioning result also deserves qualification. During data generation, obstacle-aware prompts are added to align instructions with the target behavior, and the cluttered real-world trials explicitly tell the robot to step over, sidestep or bend down. Consequently, the experiments test whether TANGO can localize and execute a requested traversal mode, but only partially test whether it can infer the correct mode autonomously from ambiguous vision and ordinary destination language. A harder benchmark would say only “go to the desk” and require the robot to decide whether to turn, duck, step or reroute. The behavior vocabulary is similarly structured around three families: stride, sidle and squat. That is enough to establish the idea, but human environments also demand stairs, irregular footholds, crawling, contact-assisted passage, movable clutter, doors and dynamic interactions with people. The reported experiments focus on static indoor geometry, so social navigation and rapidly changing obstacles remain outside the demonstrated envelope. “Zero-shot sim-to-real” should be interpreted at the navigation-policy level. TANGO itself sees no real-world navigation demonstrations, but it sits on top of SONIC, which contains a massive human-motion prior and has already been engineered for physical Unitree G1 deployment. The transfer result is still valuable: the visual policy crosses the rendering and perception gap without real navigation fine-tuning. But the whole stack is not emerging from simulation tabula rasa. Evaluation granularity is another open issue. Standard VLN success allows the robot to finish within three meters of the goal. That threshold is reasonable for comparing long-horizon route following, but it is loose for domestic robotics, where the difference between arriving at the correct side of a table and stopping across the room matters. Collision rate is episode-level in simulation, so a minor arm brush and a severe torso impact both mark the episode as collided. The real-world study counts collisions, but its 15 trials per setting remain too small to characterize rare failure modes. Scalability is also relative. PET replaces expensive robot data collection with 211 RTX PRO 6000 GPU-hours, but full VLA training adds 896 A100 GPU-hours. This is scalable compared with gathering tens of thousands of safe humanoid demonstrations, not necessarily accessible to a small laboratory. As of September 9, the paper says the pipeline, dataset, checkpoints and deployment system will be released, while the project page currently exposes the paper and videos without a public code link. Reproducibility therefore remains a promise rather than a completed artifact at this release date. A practical recipe for adapting TANGO. For practitioners, the most reusable idea is PET rather than the precise choice of VLM. If you already have a strong whole-body tracker, you can build a task-data engine around privileged simulation geometry: plan a nominal route, modify the reference motion where the full robot geometry conflicts with the scene, execute it under physics, reject failures and distill the survivors into an observation-conditioned policy. This pattern should transfer to mobile manipulators, quadrupeds carrying payloads and humanoids performing navigation while holding objects. The second lesson is to choose an action interface that exposes the coordination your task requires. A velocity command is attractive because it is compact, but it cannot communicate that the shoulders should rotate before entering a passage. Raw torque actions expose everything but force the VLA to learn dynamics and balance. Whole-body references occupy a useful middle layer: expressive enough to encode posture and gait, yet abstract enough for a robust controller to track. The third lesson is that temporal execution machinery deserves first-class treatment. TANGO’s largest ablation drop comes from removing real-time chunking, not from swapping its visual backbone or tracker. A researcher reproducing the system should therefore validate chunk alignment, inference delay, observation timestamps and controller interpolation before blaming semantic reasoning for unstable rollouts. Extensions should focus on removing privileged cues at evaluation. One direction is to train with behavior-explicit instructions but test with destination-only language, measuring whether the model selects traversal modes from geometry. Another is to add dynamic obstacle prediction and an uncertainty-aware stop or replan mechanism. Depth could be introduced as an optional teacher or fallback without abandoning RGB generalization. Finally, replacing embodiment-specific joint references with a structured, retargetable whole-body latent could preserve TANGO’s coordination benefits while making transfer across humanoid morphologies less expensive. Bottom line. TANGO’s most important idea is simple: a humanoid route is not a line on the floor. It is a moving volume through three-dimensional space, realized by an articulated body whose configuration determines whether the route is feasible. The paper does not eliminate planning, control or simulation. It uses all three aggressively to manufacture better supervision and then trains a VLA to predict the missing object between semantic navigation and physical execution: a coordinated whole-body trajectory. The success rates are not yet those of a deployable domestic robot, and the tested obstacle vocabulary remains narrow. But TANGO establishes a convincing systems blueprint. For humanoid navigation, the next waypoint may need to specify not only where the robot goes, but what shape the robot becomes on the way there.