HumanoidMimicGen: Data Engine for Loco-Manipulation

• HumanoidMimicGen is a data-generation method that transforms limited teleoperated humanoid demonstrations into thousands of adapted, collision-aware trajectories with a 0.89 success rate.
• The method employs a hybrid control strategy by combining upper-body joint-space regulation with reinforcement learning-based lower-body locomotion to ensure feasible whole-body planning.
• It integrates object-centric transformation, whole-body inverse kinematics, and GPU-accelerated collision checking to generate robust training datasets for complex loco-manipulation tasks.

HumanoidMimicGen is a data-generation method for humanoid legged loco-manipulation that starts from a small set of teleoperated source demonstrations and automatically synthesizes many more successful whole-body demonstrations in simulation by adapting contact-rich skills to new object poses, layouts, and initial states (Lin et al., 26 May 2026). Its core contribution is to combine object-centric skill adaptation with whole-body locomotion and manipulation planning, so that a humanoid can reuse a demonstrated behavior in scenes where direct replay would be unreachable, unstable, or collision-prone. In the benchmark introduced with the method, policies trained on 1,000 generated demonstrations per task achieve an average success rate of 0.89 across nine tasks, compared with 0.48 for policies trained on 100 human demonstrations and 0.33 for DexMimicGen+ (Lin et al., 26 May 2026).

1. Origins, scope, and problem setting

HumanoidMimicGen is positioned around a specific data bottleneck: imitation learning for humanoid loco-manipulation requires many demonstrations, but teleoperation for full humanoids is time-intensive and difficult because the operator must coordinate arms, legs, torso, and contact-rich interaction (Lin et al., 26 May 2026). The method is explicitly presented as a response to the limitation of existing data-generation systems for manipulators. MimicGen decomposes single-arm demonstrations into object-centric subtask segments and adapts them to new contexts (Mandlekar et al., 2023), while DexMimicGen extends that paradigm to bimanual dexterous manipulation using per-arm subtasks, synchronization, and ordering constraints (Jiang et al., 2024). HumanoidMimicGen argues that such approaches are ineffective on humanoids because their high-dimensional composite action spaces involve arms, legs, and torsos (Lin et al., 26 May 2026).

The method targets humanoid loco-manipulation rather than generic motion synthesis. Its purpose is not to learn an open-ended motion prior, nor to generate motions from language or video prompts, but to turn a handful of annotated whole-body demonstrations into a large corpus of successful training trajectories for legged humanoids. The source demonstrations are segmented into object-centric skills and annotated with precedence and coordination structure, after which the system adapts the demonstrated skills to new object poses while inserting the locomotion and upper-body motion needed to make those skills feasible in new scenes (Lin et al., 26 May 2026).

A defining characteristic is that the system is a data engine rather than the final policy. It generates successful demonstrations $D_N=\{d^1,\dots,d^N\}$, and the resulting state-observation-action trajectories are then used to train visuomotor policies by behavior cloning or related imitation-learning procedures (Lin et al., 26 May 2026).

2. Formalization and hybrid humanoid control

The robot is modeled as a bimanual humanoid with joint set

$${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$

where $J_l$ are leg joints, $J_t$ torso joints, $J_{a_l},J_{a_r}$ left and right arm joints, and $J_{h_l},J_{h_r}$ left and right hand joints. The robot configuration is

$q \in \mathbb{R}^{|{\cal J}|}.$

A source demonstration is segmented into object-centric skill demonstrations

$\psi = \langle e, f, d^\psi \rangle,$

where $e$ is an end-effector frame, $f$ is a reference object frame, and $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$0 is a contiguous subsequence of the demonstration (Lin et al., 26 May 2026).

Skill adaptation is object-relative. The paper’s adaptation rule is

$${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$1

The intended meaning is that the demonstrated end-effector action is expressed relative to the demonstration-time object frame and then re-instantiated in the current scene’s object frame. This same object-centric rule is used for skill-start targets and for actions inside the skill (Lin et al., 26 May 2026).

A central design choice is the hybrid control structure. HumanoidMimicGen does not assume that all limbs can be controlled as independent task-space effectors. Instead, it uses upper-body joint-space control for arms, hands, and torso, while the lower body is controlled through the Homie locomotion controller with base command

$${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$2

where $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$3 are planar pelvis velocities, $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$4 is yaw rate, and $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$5 is desired torso height (Lin et al., 26 May 2026). The method therefore adopts a decoupled strategy: dynamic locomotion phases are handled by the RL lower-body controller, and static manipulation phases are handled by whole-body IK and upper-body control.

This design limits the method’s scope in a precise way. It is well suited to tasks where locomotion and manipulation can be interleaved through switch states, but it is not a fully unified dynamic whole-body optimizer over simultaneous walking-and-manipulating behavior (Lin et al., 26 May 2026).

3. Skill adaptation and whole-body planning pipeline

The planning process begins from a set of annotated skills $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$6, precedence constraints $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$7, and coordination constraints. At each stage, the method selects the currently executable skills by greedily taking those with no unmet predecessors: $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$8 This gives a directed-acyclic execution structure that supports single-arm skills, dual-arm synchronized skills, and multi-stage tasks (Lin et al., 26 May 2026).

For each active skill, HumanoidMimicGen retargets the initial demonstrated end-effector pose to the current scene’s object pose, then runs a batch whole-body inverse kinematics solver to compute feasible candidate configurations $${\cal J}=J_l\cup J_t\cup J_{a_l}\cup J_{h_l}\cup J_{a_r}\cup J_{h_r},$$9 that satisfy the current end-effector targets. From each $J_l$0, it forms a “switch” configuration $J_l$1 by keeping the current upper-body joints and replacing the lower-body joints with those from $J_l$2. This switch state separates locomotion from stationary manipulation: the robot first plans and executes a lower-body trajectory to the switch configuration, then replans the upper-body motion from the achieved state to the full skill-start configuration $J_l$3, and only then replays the adapted skill demonstration (Lin et al., 26 May 2026).

The skill itself is replayed in adapted form. At each timestep of the active skill set, the system re-applies the object-centric transformation rule to the demonstrated end-effector targets, solves whole-body IK for the next upper-body waypoint, concatenates these waypoints into a skill trajectory, and executes them under upper-body manipulation control. Hand joint positions are replayed without modification, while arm and torso motions are retargeted through IK (Lin et al., 26 May 2026).

Collision checking and reachability are treated as first-class concerns. The implementation uses cuRobo with GPU-accelerated collision checking and batch IK, approximating the robot and objects by sets of spheres. For intended contact, currently colliding spheres can be shrunk until they are no longer in collision, and similar shrinking is applied near target end-effector poses for links in the same rigid connected component as the active end effector. This is a pragmatic mechanism for allowing valid manipulation contact while keeping collision tests conservative elsewhere (Lin et al., 26 May 2026).

The IK search is also biased toward minimal whole-body motion. The appendix description states that the system approximately minimizes an $J_l$4-style distance from the current configuration and progressively unlocks larger joint groups in the order

$J_l$5

thereby preferring arm-only adaptation when possible, then arm-plus-torso, then arm-plus-leg, and only finally full-body motion (Lin et al., 26 May 2026). This suggests that the planner is designed not merely for feasibility but also for economy of whole-body change.

4. Benchmark design and empirical performance

HumanoidMimicGen is evaluated on a new simulated benchmark built on robosuite and MuJoCo for a G1 humanoid (Lin et al., 26 May 2026). The benchmark spans nine tasks that vary locomotion demand, manipulation complexity, horizon length, and whether interaction is single-arm, bimanual, vertical, contact-rich, or obstacle-constrained.

Task	Interaction emphasis
Box Lift Floor	floor grasp and lift
Push Button	approach and press
Box Lift	table grasp and lift
Push Shelf Forward	bimanual contact-rich pushing
Drill Lift	table grasp and lift
Drill PnP	pick and place across tables
Box Table to Shelf	bimanual transfer into shelf
Pick Drill from Holder	extraction from holder
Drill Lift Obstacle	navigation around obstacle, then lift

In the main comparison, policies trained on a single human demonstration per task achieve an average success rate of 0.26, policies trained on 100 human demonstrations reach 0.48, DexMimicGen+ reaches 0.33, and HumanoidMimicGen reaches 0.89 (Lin et al., 26 May 2026). Task-wise success rates for HumanoidMimicGen are 0.97 on Box Lift Floor, 0.92 on Push Button, 1.00 on Box Lift, 1.00 on Push Shelf Forward, 1.00 on Drill Lift, 0.70 on Drill PnP, 0.53 on Box Table to Shelf, 1.00 on Pick Drill from Holder, and 0.87 on Drill Lift Obstacle.

The comparison with DexMimicGen+ is especially diagnostic. DexMimicGen+ uses the same single source demonstration and generates 1,000 demonstrations per task, but lacks skill reasoning, motion planning for locomotion and arm movement, and collision checking. Its average success is 0.33, which the paper uses as evidence that manipulator-style demonstration adaptation does not transfer well to humanoid loco-manipulation (Lin et al., 26 May 2026).

The strongest gains appear on long-horizon whole-body tasks. Push Shelf Forward, a 1230-step task, reaches 1.00 with HumanoidMimicGen versus 0.35 for DexMimicGen+ and 0.70 for the source-demo-only baseline. Box Lift Floor reaches 0.97 with HumanoidMimicGen versus 0.87 for DexMimicGen+ and 0.14 for source-demo-only training (Lin et al., 26 May 2026). These results indicate that the value of the method lies not only in object-pose adaptation but in the insertion of whole-body locomotion and positioning plans that make those adapted skills executable.

5. Policy learning from generated data

The generated demonstrations are used to train whole-body visuomotor policies. In the main benchmark, the strongest model is a GR00T N1.6 VLA policy fine-tuned on HumanoidMimicGen-generated data. The training setup reported for this model uses a learning rate of $J_l$6, a global batch size of 128, 25,000 training steps, and 8 NVIDIA H100 GPUs, with checkpoints evaluated every 5,000 steps over 100 episodes (Lin et al., 26 May 2026).

The paper also studies policy architecture choice on the same 1,000 generated demonstrations per task. The VLA reaches an average success of 0.89, a flow-matching policy reaches 0.86, and a diffusion policy reaches 0.51 (Lin et al., 26 May 2026). This suggests that the data-generation mechanism is not tied to one downstream learner, but also that policy class materially affects how much of the generated dataset can be exploited.

Two augmentation choices are reported as especially important during generation. First, motion noise is injected during rollout,

$J_l$7

while the original action $J_l$8 is stored as the label. Second, the robot’s initial base pose is randomly perturbed. Removing motion noise drops average success from 0.89 to 0.49, and removing initialization noise drops it to 0.51 (Lin et al., 26 May 2026). The paper interprets these perturbations as a way to expose policies to off-nominal states while preserving expert supervision.

The real-world section uses sim-and-real co-training rather than purely synthetic training. Across four real tasks—ThrowBottle, BoxToCart, PickCanister, and PickCanisterWithObstruction—policies trained only on real data average 0.51 success, while policies co-trained with HumanoidMimicGen-generated simulation data average 0.71, a gain of 0.20 or 20 percentage points (Lin et al., 26 May 2026). Task-wise improvements are 0.60 to 0.75 on ThrowBottle, 0.35 to 0.60 on BoxToCart, 0.50 to 0.75 on PickCanister, and 0.60 to 0.75 on PickCanisterWithObstruction.

6. Position within the literature and principal limitations

HumanoidMimicGen belongs to the data-generation branch of humanoid imitation research rather than to generated-video imitation, generic motion-prior learning, or pure retarget-and-track systems. MimicGen established the single-arm object-centric demonstration-generation paradigm (Mandlekar et al., 2023), and DexMimicGen generalized it to bimanual dexterous manipulation with per-arm subtask structure (Jiang et al., 2024). HumanoidMimicGen extends this lineage to legged humanoids by adding whole-body planning and a hybrid lower-body/upper-body control decomposition (Lin et al., 26 May 2026). This places it in a different category from benchmark-centric studies of scene-interaction generalization such as Mimicking-Bench (Liu et al., 2024), from generated-video humanoid-object interaction systems such as GenHOI (Bi et al., 11 Jun 2026), and from locomotion-style mimic controllers such as MuGen (Feng et al., 23 May 2026).

Its limitations are correspondingly specific. The method requires manual segmentation of demonstrations into skills and manual annotation of precedence and coordination constraints. It assumes a fixed skill vocabulary and fixed high-level sequence structure rather than open-ended task planning. Skill adaptation is rigidly object-frame-based, which the paper notes is a limitation for large intra-category shape variation or ambiguous affordances. Stability is achieved through a decoupled locomotion/manipulation architecture rather than a fully unified dynamic whole-body optimizer. The method also depends on a reliable lower-body locomotion controller, here Homie, and on simulator-specific task design, initial-state distributions, and success checks (Lin et al., 26 May 2026).

These limitations also clarify the method’s significance. HumanoidMimicGen is not a universal humanoid foundation model, but a practical data engine for whole-body loco-manipulation. Its main contribution is to show that one teleoperated source demonstration per task can be amplified into 1,000 successful whole-body demonstrations through object-centric skill adaptation, whole-body IK, collision-aware planning, and humanoid locomotion control, and that policies trained on those generated datasets can exceed both manipulator-derived generation baselines and much larger human-only datasets on the proposed benchmark (Lin et al., 26 May 2026).