Researchers built a humanoid mobility system that turns demonstrations into plans for where and when the robot should make contact, rather than asking it to replay full-body motion. In simulation, contact-only guidance reached at least 98.5% success across every reference-anchored terrain category, and the approach was deployed on a Unitree G1 with perception and control running onboard.
What did the researchers build?
The system gives a humanoid a compact plan for interacting with terrain. Instead of specifying every joint position throughout a motion, the plan describes which body parts should make contact, where those contacts should happen, and when they are expected. This keeps the important environmental interactions from a demonstration while letting the controller work out the detailed movement.
That distinction matters for whole-body movement. A robot stepping across footholds may also need to place a hand, knee, or elbow against the environment. The researchers’ planner handles nine contact-capable body parts: two feet, two knees, two hands, two elbows, and the pelvis. Its contact plans can therefore represent more than ordinary footstep sequences.
The work combines this planner with two supporting tools. A motion-retargeting pipeline adapts human movement data to the robot’s physical capabilities and records its body contacts. A LiDAR-based mapping system turns point-cloud measurements into a local terrain map. Together, these components support training on demonstrated interactions and control on terrain the robot encounters at runtime.
What were the key results?
On reference-anchored terrain, the contact-conditioned tracker achieved at least 98.5% success in every evaluated category. The comparison tested whether adding reference root and joint motion to the same contact commands improved performance. It did not produce a consistent improvement, indicating that the contact plans alone supplied enough guidance for the whole-body skills represented in this dataset.
The evaluation also covers several distinct test setups, though the reported text does not provide success percentages for all of them. These include goal-reaching terrain used for reference-free fine-tuning, terrain with added obstacles, and uneven surfaces with bumps. Eight additional compact courses combined sparse footholds, whole-body obstacles, and varying transition distances.
| Evaluation detail | Reported value | What it describes |
|---|---|---|
| Success on reference-anchored terrain | At least 98.5% in every category | Contact-conditioned tracking |
| Reference configurations | 123 | Configurations used for training and evaluation |
| Goal-reaching terrain configurations | 32 | Configurations used in Stage 3 training and evaluation |
| Curated composite courses | 8 | Separate courses combining footholds and obstacles |
| Added irregular bump amplitude | 5–15 cm | Range used to vary surface roughness |
The clearest reported comparison is therefore about representation: adding full-body motion inputs did not reliably beat contact-only guidance in the reference-anchored tests. The other terrain groups probe how well the approach handles new conditions, but the available results excerpt does not state their outcome rates. The robot was also set up to run perception and control onboard, rather than relying on an external computer.
How does the contact-planning system work?
The system builds contact plans from motion examples, then uses terrain observations to guide movement. During data preparation, the researchers retarget human motions to the humanoid and record the body parts that touch the environment. Each recorded contact has a time in the sequence and a location, so the training data captures both the order of interactions and where they occur.
At each control step, the planner produces a proposed contact for each of its nine contact-capable body parts. A target position is expressed relative to the robot’s base in a gravity-aligned coordinate frame, with yaw alignment so the plan does not depend on the robot’s absolute position or heading. A binary flag indicates whether a body part is expected to establish a new contact within the planning horizon. The resulting plan leaves the controller room to choose the detailed joint motion.
Terrain sensing is handled by a neural mapping pipeline designed for the robot’s head-mounted LiDAR. It divides point-cloud data into local grids and predicts elevations. For each grid, the network can keep the measured height, use a neighborhood average, or leave the grid empty. This lets it handle sparse LiDAR returns; its predictions also provide an uncertainty estimate that helps the map account for noisy readings.
Training proceeds through stages, including reference-guided tracking and reference-free goal-reaching fine-tuning. The system also randomizes physical properties such as friction, mass, and joint behavior, and adds noise to observations. At deployment, the Unitree G1’s LiDAR, state estimator, mapping pipeline, and learned policy run concurrently on its onboard NVIDIA Jetson Orin.
Why does this matter for robotics?
Many robot controllers face a choice between replaying a detailed motion and learning movement directly through trial and error. A detailed reference can help reproduce a known action, but it can be hard to adapt when the terrain differs from the demonstration. Learning from scratch can be flexible, but whole-body interactions are difficult to teach through reward design alone.
Contact plans offer a middle ground: demonstrations identify the important places and moments of interaction, while the controller retains flexibility over how to move between them. That could make it easier to reuse learned behaviors across changing terrain, including tasks where hands or knees provide support as well as feet. It also creates a shared representation for actions such as walking, climbing, crawling, and recovery.
The onboard setup is relevant to teams evaluating humanoids for environments where external sensing or compute is inconvenient. It does not, by itself, establish readiness for every deployment task, but it shows how sensing, mapping, and learned control can be integrated on the robot. Teams comparing platforms can browse humanoid robots on Robot Overflow, or explore used robots for sale.
What are the limitations and open questions?
The current sensing and contact representations focus mainly on terrain-based, or 2.5-D, interactions. The mapping system captures fine terrain features such as ladder rungs, but the approach does not yet provide the richer 3D representation needed for actions such as hanging, overhead contact, or manipulation of surrounding objects.
The system also addresses local goal-conditioned locomotion, not global navigation. It cannot distinguish motions that share the same contact sequence but require different body movement: a backflip and a vertical jump are given as examples. Aerial actions without distinct contact patterns are therefore a poor fit for the current representation.
Broader deployment will also require evidence beyond local terrain tasks. The paper identifies vision-based and general 3D scene representations, semantic task inputs, and higher-level planning as future directions. These additions could help choose goals and distinguish intended behaviors when contact information alone is ambiguous.
What are the most common questions about the system?
Does the robot need a full-body motion script at runtime? No. The method uses contact plans to guide movement rather than requiring a full reference trajectory during deployment.
Which body parts can the planner include? It plans contacts for the feet, knees, hands, elbows, and pelvis.
What robot hardware was used for deployment? The system ran on a Unitree G1, with perception, mapping, state estimation, and policy inference onboard its NVIDIA Jetson Orin.
Does the reported 98.5% success apply to every terrain test? No. That figure applies to every reference-anchored terrain category. The available results excerpt does not give success rates for the other terrain groups.
What should robotics teams take away?
The central idea is to preserve the useful structure in a movement demonstration without locking the robot into a full-body script. Contact plans provide a compact way to guide whole-body mobility, and the reported results show they can be sufficient for strong tracking performance on the tested reference-anchored terrain. Further evidence is needed to establish performance across broader 3D interactions and navigation tasks.