Source-linked AI summary
Humanoid Mobility via Attention-based Contacts
Junzhe He, Kehan Wen, Chong Zhang, Clemens Schwarke, Zewei Zhang, Kaixian Qu, Siyuan Hu, Matthias Müller, Vladlen Koltun, Marco Hutter
TL;DR
Whole-body mobility over complex terrain requires choosing coordinated body–environment contacts, while reward-driven exploration and reference tracking have limitations. The paper uses an attention-based contact planner and contact-conditioned tracker, followed by RL fine-tuning; in simulation, fine-tuning improves performance on goal-reaching and unseen terrains.
Problem
Complex terrain requires selecting the spatial, temporal, and anatomical distribution of contacts, while reward shaping alone makes whole-body coordination challenging.
Source: INTRODUCTION · p. 1Method
An attention-based planner predicts future body–environment contacts from terrain, proprioception, and goals, and a contact-conditioned tracker executes them, with motion data used as training guidance.
Source: INTRODUCTION · p. 3Results
The Stage 3 attention-based policy retains high success on reference-anchored terrains and improves success and contact placement accuracy in every goal-reaching and unseen category, with a 54.8 percentage-point success increase on sparse whole-body terrain.
Source: §C.3 · p. 11Takeaways & Limitations
Selective terrain attention and explicit contact prediction support transfer of demonstrated contact strategies to new terrain geometries and goal-conditioned behaviors.
Source: DISCUSSION · p. 12The system targets primarily terrain-centric 2.5-D interactions and local goal-conditioned locomotion, and cannot distinguish aerial motions with the same contact sequence.
Source: DISCUSSION · p. 12
Abstract
from arXiv · showhide
Humanoid mobility over complex terrain requires accurate foot placement and coordinated whole-body motion capable of robust adaptation. Reinforcement learning (RL) policies trained without motion references generalize well across terrains, yet eliciting effective whole-body coordination through reward design and exploration alone remains challenging. In contrast, reference-based methods yield natural motions but struggle to adapt beyond the predefined trajectories in their training data. We introduce a learning framework that bridges these two approaches by using motion data strictly as a training prior. An explicit, self-supervised, attention-based contact planner predicts future body-environment contacts, while a contact-conditioned tracker translates these plans into coordinated whole-body motion. Together, they establish contact as an interface between sensory observations and control actions. This abstraction allows the policy to internalize the contact-aware structure of locomotion rather than directly replicating kinematics. Consequently, a single policy can exploit data-driven motion priors for natural whole-body coordination while retaining the flexibility to discover environment-dependent behaviors through reinforcement learning. Building on the learned contact planner and tracker, a subsequent fine-tuning stage bootstraps their acquired capabilities to explore environment variations beyond those covered by the original motion data. Operating entirely from onboard observations without any reference inputs at deployment, this framework enables versatile humanoid mobility in challenging environments.
INTRODUCTION
Complex-terrain humanoid mobility requires selecting where and which body parts interact while coordinating whole-body motion, but reward-driven exploration and reference-based tracking each have limitations. The framework uses contact plans as an interface between terrain perception and control, using motion data as training guidance while retaining freedom to adapt at deployment.
INTRODUCTION
- Contacts preserve task-relevant environmental interactions while leaving the policy free to choose among whole-body motions that realize them.Unlike full-body references, contact targets can constrain sparse-support interactions without prescribing the motion between contacts.Source: INTRODUCTION · pp. 3, 5
- The policy requires no reference input at deployment and uses onboard observations to adapt contact strategies to terrain and goals.The framework uses motion data as training-time guidance rather than requiring motion retrieval or a motion generator at deployment.Source: INTRODUCTION · pp. 3, 5
- The contact planner predicts explicit terrain-map contacts from attention conditioned on robot proprioception and the user-specified goal.It selects locations for feet, hands, and other body contacts as the robot moves, reducing the planning problem's dimensionality relative to full-trajectory prediction.Source: INTRODUCTION · p. 3
- The framework uses an explicit, self-supervised attention-based contact representation to connect perception with whole-body control.Motion data guides discovery of useful body–environment interactions rather than serving as full trajectories to reproduce.Source: INTRODUCTION · pp. 3, 5
- The framework combines a contact-conditioned tracking pipeline, a nine-body-part attention-based contact planner, RL fine-tuning, and LiDAR mapping for fine-grained geometry.The listed planner body parts are hands, elbows, knees, feet, and pelvis; the mapping framework captures small structures such as ladder rungs.Source: INTRODUCTION · p. 5
- Training proceeds through contact tracking, contact planning, and RL fine-tuning, progressively reducing reliance on motion data while retaining demonstrated coordination.Fine-tuning uses terrains outside the demonstration distribution and successful rollouts to provide additional planner supervision.Source: Figure 2 · p. 4, INTRODUCTION · p. 5, Figure 3 · p. 5
A. Agile and generalized whole-body locomotion
Real-robot experiments show a single policy producing varied whole-body behaviors and traversing unseen, composite terrain configurations without deployment reference motions.
A. Agile and generalized whole-body locomotion
- The Unitree G1 policy performs diverse indoor and outdoor whole-body behaviors, including stair, ladder, and box climbing, sitting, and standing from the ground.At deployment, it operates without reference motions and adapts contact strategy to perceived terrain.Source: A. Agile and generalized whole-body… · p. 5
- On the same box, different goal positions and orientations elicit sitting, stepping, climbing, crawling, or sideways climbing behaviors.The policy jointly interprets geometry and goal commands rather than mapping each terrain category to a fixed motion primitive.Source: A. Agile and generalized whole-body… · p. 6
- The policy traverses unseen composite terrains and connects skills without manually designed transitions, explicit skill switching, or terrain-specific motion augmentation.For unstructured terrains, the experiments use manually specified intermediate waypoints; global navigation is outside the work's scope.Source: A. Agile and generalized whole-body… · p. 6
B. Failure Recovery
Real-robot trials report recovery from unsuccessful stand-up attempts, falls, and getting stuck during ladder climbing.
B. Failure Recovery
- The policy repeatedly attempts recovery after failed stand-up, jump-down, perception-related fall, and ladder-climbing episodes.These recovery trajectories are not explicitly represented in the original motion dataset.Source: B. Failure Recovery · p. 6
C. Simulation-based Evaluations
The simulation evaluations test tracker guidance, transfer beyond reference terrains, and the roles of fine-tuning and attention-based planning across defined terrain groups.
C. Simulation-based Evaluations
- The evaluation separates reference-anchored terrains, goal-reaching terrains, and unseen terrains to test retention, zero-shot transfer, and generalization properties.Unseen terrains vary obstacle composition, distractors, and surface roughness to probe skill transitions, rejection of irrelevant geometry, and irregular-support traversal.Source: C. Simulation-based Evaluations · p. 6
- The terrain taxonomy distinguishes dense or sparse support and foot-only or whole-body support requirements.Sparse foot-only terrains demand precise placement, while whole-body categories distinguish broad additional support from discrete multi-link contacts.Source: C. Simulation-based Evaluations · p. 6
- Category comparisons pool 3,000 episodes per method and category across three seeds; terrain-specific comparisons use 1,000 episodes per method and radar axis.Success is generally reaching within 0.5 m of the active goal, with specified exceptions for postural transitions and curated courses; sensor noise and several other factors are disabled.Source: C. Simulation-based Evaluations · p. 6
C.1. Ablation study on tracker design
The tracker ablation finds that contact commands alone guide whole-body skills without full-body reference kinematics as policy inputs.
C.1. Ablation study on tracker design
- Contacts alone achieve at least 98.5% success in every category, with no consistent improvement from adding reference kinematics.Both tracker variants receive the same reference contact commands and retain reference-motion guidance during training.Source: §C.1 · p. 6
- The comparison adds reference root and joint kinematics to a tracker that already receives the same reference contact commands.Both variants retain reference-motion guidance during training.Source: §C.1 · p. 6
- Contact plans alone suffice to guide the dataset’s whole-body skills without high-dimensional full-body reference kinematics as policy inputs.Source: §C.1 · p. 6
C.2. Ablation study on planner architecture and generalization
The section compares attention and MLP contact planners and illustrates terrain-sequence compositions performed by one generalist policy.
C.2. Ablation study on planner architecture and generalization
- The planner comparison tests attention and MLP systems predicting per-body contact plans over a flattened height scan.Source: §C.2 · p. 6
- The two planner variants have comparable parameter counts.Source: §C.2 · p. 6
- One generalist policy performs sequences combining walking, climbing, vaulting, beam traversal, and jumping down.The examples include transitions among ladders, stairs, stepping stones, boxes, and beams.Source: Figure 4 · p. 7
A. Diverse motions conditioned on different goal commands on the same type of terrain
The examples show goal-dependent interaction choices on a given terrain, alongside planner comparisons and generalization across terrain conditions.
A. Diverse motions conditioned on different goal commands on the same type of terrain
- On the same terrain, the robot selects different interaction strategies depending on the desired goal position and orientation.Figure 5 also depicts generalization to composite and unseen terrains.Source: Figure 5 · p. 8
- The generalist controller exhibits recovery behaviors including repeated stand-up attempts on uneven ground, recovery after falls, and recovery from ladder-related failures.Source: Figure 6 · p. 9
- After Stage 3 fine-tuning, attention outperforms the MLP across goal-reaching terrain categories, with clear gains on unseen composition and distractors.The largest advantages occur on sparse whole-body terrain and with distractors.Source: A. Diverse motions conditioned on… · p. 11
- With Stage-2 planners frozen, attention-based planner–tracker pairings achieve higher mean success in all three terrain groups.The comparison uses each planner with its corresponding adapted Stage-3 tracker.Source: A. Diverse motions conditioned on… · p. 11
C.3. Ablation study on RL fine-tuning
Fine-tuning improves performance on goal-reaching and unseen terrains, while the comparison with MGMT reports higher success across terrain categories and composite courses.
C.3. Ablation study on RL fine-tuning
- Stage 3 fine-tuning improves success and contact-placement accuracy in every goal-reaching and unseen terrain category while retaining high success on reference-anchored terrains.The largest success gain is 54.8 percentage points on sparse whole-body terrain.Source: §C.3 · p. 11
- Fine-tuning gains extend beyond its training conditions, with composition gains indicating better transitions between familiar interactions in new obstacle sequences.Improvements also occur with distractors and roughness.Source: §C.3 · p. 11
C.4. Quantitative comparison with MGMT
- Our method achieves higher success than MGMT in all eleven terrain categories and all eight composite courses, averaging 92.5% versus 58.8% on the courses.The courses combine sparse footholds with whole-body obstacles and test transitions between support and coordination demands.Source: §C.4 · p. 11
C.5. Qualitative benchmark with other whole-body control methods
- The qualitative benchmark is not a direct quantitative comparison because methods differ in data, platforms, terrain definitions, perception, and training pipelines.N/A means a capability was not demonstrated or evaluated, not that a method cannot perform it; arbitrary terrain transitions for LLP remain unclear from reported results.Source: §C.5 · p. 11
DISCUSSION
The paper frames contact as a shared interface between perception and whole-body control: contacts specify where, when, and with which body part to interact without fixing the full motion. The authors identify limits in the current representation and outline extensions for broader environments and tasks.
DISCUSSION
- Contact specifies where, when, and with which body part to interact, leaving reinforcement learning room to explore whole-body behaviors on new terrain.The authors report that a single policy can compose and transition between behaviors without reference trajectories, skill labels, or manually designed transitions at deployment.Source: DISCUSSION · p. 11
- Selective terrain attention and explicit contact prediction help transfer demonstrated contact strategies to new geometries and unseen terrain compositions.Goal conditioning allows different contact strategies on the same terrain for different desired positions and orientations.Source: DISCUSSION · pp. 11–12
- The current representation primarily targets terrain-centric 2.5-D interactions and local goal-conditioned locomotion, not general 3-D interactions or global navigation.Overhead contacts, hanging, and manipulation would require richer 3-D perception and contact representations.Source: DISCUSSION · p. 12
- Aerial motions with indistinguishable contact sequences, such as a backflip and vertical jump, cannot be distinguished by the current contact representation.The paper suggests semantic task inputs as a possible way to specify intended motion when contact sequences are shared.Source: DISCUSSION · p. 12
- The authors propose extending the interface to vision-based, general 3-D scenes and combining it with semantic or global planners for larger environments.They frame walking, climbing, crawling, recovery, and manipulation as sequences of body–environment interactions.Source: DISCUSSION · p. 12
MATERIALS AND METHODS
The method uses contact as a compact interface between demonstrated motion, terrain perception, and whole-body control. A tracker learns to realize contacts, an attention-based planner predicts them, and staged training gradually shifts from reference guidance to terrain-specific reinforcement learning.
MATERIALS AND METHODS
- Contacts retain the spatial, temporal, and anatomical structure of demonstrated interactions without transferring complete trajectories, separating which interactions occur from how the robot realizes them.At deployment, reference motions are removed and reinforcement learning synthesizes motion between contact targets.Source: Motivation · p. 12
- Contact-aware retargeting transfers human behaviors to the robot while extracting contact sequences and locations alongside robot kinematics.The retargeting optimization accounts for robot limits, ground constraints, and environmental geometry to maintain physically consistent interactions.Source: Contact-Informed Whole-Body Tracker · p. 12
- The tracker receives proprioception, a local height scan, nine contact targets, and a goal command, then outputs joint-position offsets for control.The paper reports that contact commands suffice to execute demonstrated skills without reference kinematics as tracker inputs.Source: Contact-Informed Whole-Body Tracker · p. 13
- The planner predicts target positions and occupancy for nine contact-active bodies, enabling plans for multi-contact behavior beyond footstep sequences.Targets are expressed relative to the robot base in a gravity-aligned, yaw-rotated frame.Source: Attention-based Whole-Body Contact… · p. 13
B. Neural Network Design
- The attention planner uses proprioception, goal commands, and terrain points to select surface-based contact targets and predict contact occupancy.For quantitative evaluations, each retained target coincides with an observed terrain point.Source: Attention-based Whole-Body Contact… · p. 13
MATERIALS AND METHODS
- Training first teaches the tracker with reference-derived contacts, then jointly trains the planner and tracker using successful rollout contacts as supervision.The first stage supplies ground-truth contacts; the second uses hindsight labels from successful rollouts and gradually transfers contact guidance to planner predictions.Source: Three-Stage Training Pipeline · pp. 13–14
- A third-stage bootstrapping loop fine-tunes on procedural goal-reaching terrains without reference motions, using successful executions to refine planner supervision.Reference-trajectory training remains in the mix to preserve previously learned tracking behaviors.Source: Three-Stage Training Pipeline · p. 14
B. Neural Network Design
- The LiDAR mapping pipeline predicts grid elevations and uncertainty from local scans, then fuses predictions into a global map using odometry.It classifies each grid as measured height, neighborhood average, or empty; disagreement among modes and logits estimates uncertainty.Source: LiDAR-based Neural Mapping · p. 14, Figure 8 · p. 15
Training
The authors train the tracker with PPO and train the contact planner from tracker rollouts across three stages. The resulting policy is deployed on a Unitree G1 with perception and control running onboard.
Training
- The three training stages take approximately 1 day, 1 day and 21 hours, and 2 days and 6 hours, respectively, using 2–4 RTX 4090 GPUs.Stage 1 uses 3 GPUs, Stage 2 uses 2, and Stage 3 uses 4.Source: Training · p. 16
- Domain randomization varies physical parameters at startup and adds noise to actor observations throughout all three training stages.Height scans also receive per-step Gaussian noise and a per-episode elevation offset.Source: Training · p. 16
Deployment
- The trained policy runs on the Unitree G1 using onboard LiDAR, state estimation, neural mapping, and policy inference on the Jetson Orin.The deployment uses no external computation or sensing.Source: Deployment · p. 16
Training Details
Training details specify reference and procedural terrain curricula, policy and planner architectures, and the replay and optimization procedures used to train them.
Training Details
- The tracker and value networks use three hidden layers of 512, 256, and 128 units, while the attention and MLP planners contain 786,993 and 833,322 parameters, respectively.The contact-conditioned tracker receives 468 features and outputs commands for 23 joints.Source: Network architectures · p. 20
- Planner updates use successful-episode replay, with ten optimizer updates per PPO iteration after at least 1,000 replay entries are available.Batches draw 4,096 replay entries, with sampling balanced across terrain types and tiles when buffers are partitioned.Source: Planner optimization and hindsight… · p. 22
- During Stage 2, contact-input selection and contact-reward weighting follow separate linear schedules, with the reward transition starting later.The input schedule selects reference or predicted commands per environment and control step rather than interpolating contact coordinates.Source: Contact and reward annealing · p. 22
- PPO performs five epochs with eight minibatches per epoch, and Stage 3 normalizes advantages separately for reference-following and goal-reaching samples.Goal-reaching samples are excluded from the actor's PPO surrogate loss for the first 500 iterations while training the new critic.Source: PPO parameters · p. 23
Evaluation Details
Evaluation covers reference-anchored, goal-reaching, and unseen terrains, plus a separate set of curated composite courses. The reported course success rates use 1,000 episodes per method and course.
Evaluation Details
- Reference-anchored evaluation used the same 123 reference configurations as training.Source: Evaluation Details · p. 23
- The same 32 terrain configurations used in Stage 3 training had difficulty levels sampled evenly from the training range; the cited passage does not identify the evaluation task.Source: Evaluation Details · p. 23
- Unseen-terrain evaluation included irregular surface bumps with maximum amplitudes from 5 to 15 cm.Source: Evaluation Details · p. 23
- Eight curated composite courses were evaluated separately from the unseen-terrain benchmark.Each course combines sparse footholds with whole-body obstacles and varied transition distances.Source: Evaluation Details · p. 23, Figure S1 · p. 25
- The PPO settings table counts batch sizes in transitions and does not report Stage-1 distributed totals.Source: Table S6 · p. 24
- Success rates for each curated course were measured over 1,000 episodes per method, with a 0.5 m horizontal goal tolerance.Source: Figure S1 · p. 25