Robots Need Both Foresight and Feedback Before They Commit to Uncertain Physics
A terrain predictor and a rope-probing system point to a practical split in robot autonomy: vision can flag likely trouble before contact, while a small physical action can reveal properties that images alone cannot settle.
By Jonas Vale · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human field experience or credentials.
Key points
- TACOS turns signals from legged state estimation into visual supervision for predicting whether a foothold may be unstable before a robot steps there; its reported evaluations separate contact-scoring results from visual-prediction results.
Sources: S1
- CMU’s Wiggle and Go uses a brief rope motion and camera observation to estimate physical properties before planning manipulation, addressing behavior that the researchers say cannot be determined from appearance alone.
Sources: S2
- The systems are complementary rather than interchangeable: one seeks to reduce risk through learned visual anticipation, while the other deliberately gathers physical evidence when latent material behavior is central to the task.
The decision is not simply whether to perceive, but when to touch
Legged locomotion and rope manipulation create different versions of the same operational problem: a robot must act while important physical behavior remains uncertain. TACOS is designed to make a visual assessment before a footstep, learning from a footstep-instability score derived from the residual of a factor graph used in legged state estimation. The score is projected into images that preceded contact, producing supervision for a dense visual predictor. Its reported forest demonstration describes a robot handling non-geometric terrain properties including mud and vegetation while stepping around fallen branches.
Sources: S1
The Wiggle and Go approach begins from a different constraint. Carnegie Mellon researchers describe ropes and cables as objects whose bending, twisting and stretching make their movement difficult to predict. They say visual appearance alone cannot reveal key differences such as length, weight and stiffness. Instead of trying to infer all of that passively, the robot first gives the object a brief wiggle, an overhead camera records the traveling wave, and computer-vision tracking feeds a model that estimates physical properties for subsequent planning.
Sources: S2
What the terrain result actually measures
TACOS reports a stronger empirical record in the material supplied, but its reported gains belong to distinct tests and should not be treated as one combined operating result. On 1716 manually annotated contacts, both its hindsight-based hard score and immediately available causal soft score outperformed the evaluated alternatives for instability detection and severity ranking. The paper reports a 28 % improvement in pooled severity ranking over the strongest baseline in that evaluation.
Sources: S1
A separate evaluation used three outdoor test runs to train visual predictors. There, the TACOS-derived metric reportedly produced the most accurate predictor, with rank correlation improving by 58 % over the strongest alternative and by a further 25 % when onboard context was added. The supplied abstract does not identify the content of that onboard context or establish that the forest behavior, the contact-label evaluation, and the visual-predictor comparison occurred under identical conditions. Those distinctions matter for anyone deciding whether the result is ready to transfer to another robot, terrain type, or sensing stack.
Sources: S1
The practical value is that a mobility system could use a pre-contact visual estimate to alter where it steps before loading the ground. Yet TACOS’s hard score uses hindsight, while the paper presents the soft score as causal and available immediately. That makes the causal variant the closer match to a real-time decision, even though both variants are reported to beat the evaluated alternatives in the contact analysis.
Sources: S1
Sources: S1
A wiggle treats uncertainty as something to measure
Wiggle and Go places its information-gathering action before the main manipulation rather than learning the task through repeated failed attempts. The CMU account says the model was trained entirely in simulation, then uses the observed motion from the initial wiggle to estimate properties and improve real-world movement predictions. The robot can then plan a motion toward a goal or a desired configuration using those estimates.
Sources: S2
That sequencing is significant for deployment settings where a failed first action is costly. CMU gives rescue-rope throwing as an example, while also identifying manufacturing, construction, healthcare, environmental monitoring and search and rescue as prospective areas for more reliable cable and rope handling. These are application possibilities described by the institution, not evidence in the supplied material that the system has been deployed in each setting.
Sources: S2
The infrastructure dependency is also clearer than it may first appear. The method needs a robot that can safely excite the object, an overhead camera able to observe the response, vision algorithms that can track it, and a planning system able to use the estimated properties. A robot that has a good manipulation policy but lacks a usable view of the wave, or lacks room for the probing motion, may not obtain the information on which the plan depends.
Sources: S2
Sources: S2
Inference: choose passive prediction when contact is the hazard, probing when ambiguity is the hazard
The cross-source lesson is not that every robot should add a wiggle before acting. The evidence supports a narrower inference: passive visual prediction is especially attractive when the system can learn useful pre-action cues from prior physical outcomes, as TACOS attempts for footholds. Active probing becomes more compelling when the decisive variables are not visually identifiable, as CMU says is true for rope properties. The choice is therefore a trade-off between preserving momentum and deliberately spending an action to reduce uncertainty.
That distinction also exposes a safety limit. A visual terrain predictor may be valuable because it avoids committing weight to questionable ground, but it depends on whether visible cues and available onboard context generalize to the next environment. A rope probe can produce individualized evidence about the object in front of the robot, but it introduces requirements for observation quality, physical clearance and a probe that does not itself create unacceptable risk. Neither supplied account establishes how either method performs when those operating assumptions fail.
For system designers, the more useful architecture may be a decision layer that distinguishes uncertainty types: predict from vision when prior contact data make appearance informative; probe when object-specific dynamics must be estimated; and decline or hand off when neither route can produce a sufficiently dependable plan. This is an inference from the different designs, not a reported feature of either system.
What would change the assessment
The terrain case would be stronger for operational adoption with evidence that separates the causal soft score’s real-time behavior from hindsight-assisted evaluation across additional robots, terrain conditions and sensing configurations. It would also help to know whether the reported gain from onboard context persists when that context is unavailable or degraded. The supplied TACOS material is an abstract for a paper submitted to ICRA 2027, so it provides reported results but only limited methodological and deployment detail.
Sources: S1
The rope case would be easier to compare with other manipulation methods with measured outcomes for planning accuracy, task completion, failure modes and the effect of camera or tracking degradation. The supplied CMU account says the work was accepted to the 2026 Conference on Robotic Learning and describes real-world prediction improvement after simulation training, but it does not provide those comparative measurements. Until such evidence is available, the strongest shared conclusion is architectural: reliable physical autonomy may require robots to know when seeing is enough and when the environment must answer back.
Sources: S2
Why it matters
These developments shift attention from a single “better model” toward the operational loop around it: what the robot can infer before commitment, what sensing and physical interaction it needs to resolve uncertainty, and when an unrepeatable action demands a more conservative plan. Terrain and deformable-object tasks may look separate, but both show that dependable deployment depends on matching information collection to the physical risk of the next move.
Sources
- TACOS: Terrain-Aware COntact Supervision for Visual Footstep Planning — arXiv Robotics ·
- A Little Wiggle Helps Robots Learn the Ropes — Carnegie Mellon Robotics Institute ·