HOTICE: Whole-Body Humanoid Object Transportation in Cluttered Environments

Method in Detail

Two fields: Standard and Object-Centric Guidance

The Standard Guidance Field (SGF) follows HumanoidPF: an attractive field proportional to geodesic distance and a repulsive field piecewise in the signed distance,

$$\mathbf{U}_{\text{att}}(\mathbf{x}) = \eta\,\mathrm{D}_{\text{geo}}(\mathbf{x},\mathbf{g})$$

$$\mathbf{U}_{\text{rep}}(\mathbf{x}) = \begin{cases} \tfrac{1}{2}\xi\left(\tfrac{1}{\mathrm{sdf}(\mathbf{x})}-\tfrac{1}{d_{0}}\right)^{2}, & \mathrm{sdf}(\mathbf{x}) \leq d_{0} \\ 0, & \mathrm{sdf}(\mathbf{x}) > d_{0} \end{cases}$$

where $\mathrm{D}_{\text{geo}}$ is the shortest obstacle-free path length and $\mathrm{sdf}$ the signed distance to $\mathscr{C}$. The guidance field is the negative gradient of the summed potential,

$$\mathbf{F}_{\text{standard}}(\mathbf{x}) = -\nabla_{\mathbf{x}}\Big(\mathbf{U}_{\text{att}}(\mathbf{x}) + \mathbf{U}_{\text{rep}}(\mathbf{x})\Big)$$

with $d_{0}=0.2$ m, $\eta=0.6$, $\xi=1\times 10^{-3}$ inherited from HumanoidPF. SGF vectors are queried at head, shoulders, torso, pelvis, knees, and feet; for head, pelvis, and feet the reward maximizes the cosine similarity between $\mathbf{F}_{\text{standard}}(\mathbf{x})$ and the part's Cartesian velocity $\mathbf{v}$.

The Object-Centric Guidance Field (OGF) is the first key change: inflate obstacles by margin $m=8$ cm into $\mathscr{C}_{\text{inflated}}$, then compute the same attractive, repulsive, and guidance quantities with respect to the inflated set. Two motivations drive the margin. The object is not directly actuated by joints, so the margin buys robustness to sensor noise and control delay. And object avoidance must be more anticipatory: feet can step over a ground bar just as they reach it, whereas a heavy or bulky object becomes very hard to reposition once close to an obstacle, so it must start yielding earlier. Figure 2 shows the orange OGF bending away sooner than the blue SGF near the obstacle.

OGF vectors are queried at object keypoints (the eight corners of a box) and at the hands that control it. Because the state-based policy is a lightweight MLP, keypoint count does not hurt latency and the granularity stays controllable — the hook that later lets shape changes live entirely in keypoint placement.

The OGF steering reward is a relative quantity: maximize the cosine similarity between $\left[\mathbf{F}_{\text{object}}(\mathbf{x})-\mathbf{c}\right]$ and $\left[\mathbf{v}-\mathbf{c}\right]$, where $\mathbf{c}$ aggregates the overall navigation command from the SGF vectors of head, pelvis, and feet. Subtracting $\mathbf{c}$ removes the motion locomotion already produces, so the reward credits only the residual, obstacle-driven adjustments the hands are actually responsible for — otherwise the hands would free-ride on walking displacement.

Figure 2: SGF (blue) versus OGF (orange) in one scene. OGF is computed against the inflated obstacle (beige) and bends earlier at its boundary, granting the payload headroom (paper Figure 3).

Dual agents: splitting 20 dimensions into 8 + 12

Following FALCON's spirit, HOTICE refuses to treat the whole policy as one learner: an upper-body agent controls 8 arm joints and a lower-body agent controls 12 leg joints (waist and wrist joints are dropped as unnecessary). Each agent keeps its own actor and privileged critic, yet both read the same observation and share termination conditions — coupling happens through the environment rather than a monolithic network. The actors emit $\mathbf{a}_{\text{low}}\in\mathbb{R}^{12}$ and $\mathbf{a}_{\text{up}}\in\mathbb{R}^{8}$, concatenated into the 20D action fed to a PD controller.

Reward terms split into three sets: $\mathcal{L}$ (locomotion, balance), $\mathcal{U}$ (object reaching, grasp maintenance), and $\mathcal{S}$ (object lifting plus HOD-PF rewards whose values depend jointly on arm articulation and torso pose). The two reward streams are

$$r_{\text{low}}=\sum_{i\in\mathcal{L}\cup\mathcal{S}}w_{i}r_{i},\qquad r_{\text{up}}=\sum_{i\in\mathcal{U}\cup\mathcal{S}}w_{i}r_{i}$$

Each agent forms its own advantage $\hat{A}_{\text{low}}$, $\hat{A}_{\text{up}}$ from its own stream and value function, and the two clipped PPO objectives are optimized jointly:

$$\mathcal{J}=\mathcal{J}^{\text{PPO}}_{\text{low}}\big(\hat{A}_{\text{low}}\big)+\mathcal{J}^{\text{PPO}}_{\text{up}}\big(\hat{A}_{\text{up}}\big)$$

Because the shared set $\mathcal{S}$ enters both streams, both agents remain jointly accountable for the carry, while locomotion and grasping signals no longer compete inside a single advantage estimate — precisely the pathology that makes the single-agent variant unstable in extreme postures.

Specialist to generalist: 75 teachers into one student

Scene-specific teachers train in parallel with privileged states and without any scene-specific reward tuning — the decoupled potential fields already serve as a generalizable representation and reward structure. Teachers are distilled into a single student via DAgger, the student is refined with PPO under the same dual-agent framework, and sensor noise plus control delays are simulated during its training. Everything runs in MuJoCo Playground. The main experiment trains teachers for 400M timesteps on 20 scenes; the generalization experiment scales to 75 teachers on 75 scenes before testing the generalist on 50 unseen environments.

Shape generalization lives in the keypoints

For cylinders the field is queried at four evenly spaced points on the top rim and four on the bottom rim; for spheres, at eight fixed world-frame octant directions at the sphere's radius. Every shape uses exactly eight keypoints, so observation layout, network architecture, and action space stay unchanged; only minimal shape-specific reward adaptations (hand placement, object uprightness) are added while all other reward terms, weights, and hyperparameters are inherited from the box policy. Shape information collapses into one question: where do the eight queries sit?

Real-robot deployment: SLAM map plus AprilTag box pose

On hardware, the G1 first walks the empty environment to reconstruct obstacles via SLAM. At inference, HOD-PF guidance vectors come from real-time localization and joint angles, and the box pose is estimated from AprilTags on the box plus the head camera; the pickup and transportation policies both run at 50 Hz. Live LiDAR reconstruction is possible in principle, but the G1's single LiDAR has its field of view occluded by the carried box — a hardware limitation, the authors stress, not a method limitation.

flowchart TD
  SCN["Cluttered scene: side, overhead, ground obstacles"] --> SDF["Obstacle signed distance field"]
  SDF --> SGF["Standard Guidance Field queried at head, shoulders, torso, pelvis, knees, feet"]
  SDF --> INF["Obstacle inflation by 8 cm margin"]
  INF --> OGF["Object-Centric Guidance Field queried at 8 object keypoints and both hands"]
  SGF --> OBS["Shared observation: guidance vectors, proprioception, object state"]
  OGF --> OBS
  OBS --> UP["Upper-body actor and privileged critic: 8 arm joints"]
  OBS --> LO["Lower-body actor and privileged critic: 12 leg joints"]
  UP --> RUP["Reward stream r_up from U set plus S set"]
  LO --> RLO["Reward stream r_low from L set plus S set"]
  RUP --> JOINT["Joint PPO objective J; shared S set keeps the body coupled"]
  RLO --> JOINT
  JOINT --> PD["20D action through PD controller to torques"]
  PD --> TEA["Scene-specialist privileged teachers trained in parallel, no per-scene tuning"]
  TEA --> DAG["DAgger distillation plus PPO refinement with noise and delays"]
  DAG --> STU["Generalist student policy"]
  STU --> REAL["Real Unitree G1: SLAM map, AprilTag box pose, 50 Hz"]
  

Figure 3: Method overview. Top left: the two decoupled guidance fields. Top right: the dual-agent architecture with shared observations and shared rewards. Bottom: teacher-student distillation into the deployed policy (paper Figure 2).

Experiments

With no prior method addressing this setting, the main comparison is self-ablation: SGF only (a naive application of HumanoidPF to all robot and object keypoints), single-agent with HOD-PF, and single-agent with SGF. All four methods train teachers for 400M timesteps on 20 scenes and distill them into students; box mass and size are randomized; evaluation injects input noises, delays, and perturbations; 20 scenes × 1,000 rollouts give 20,000 trials. Metrics are success rate (reach within 0.1 m of the target within 5 s without falling, dropping, or colliding) and mean distance to goal (mDist).

Table 1: Simulation main results over 20,000 trials in 20 scenes (paper Table I).

Reading the grid contribution by contribution: adding OGF under a single agent lifts SR from 58.7 to 65.7 (+7.0); the dual-agent architecture with SGF alone already reaches 77.1 (+18.4 over single-agent SGF); only both together reach 88.5 (+11.4 over dual-agent SGF). mDist falls from 0.296 m to 0.130 m, a 56% reduction. Figure 4(A) shows the two failure modes side by side: Dual-Agent & SGF clips the obstacle with hand and box — right architecture, coupled field — while Single-Agent & HOD-PF loses balance and drops the box stepping over a ground bar — right field, fragile architecture. The contributions are complementary, not redundant.

Figure 4: (A) Ablation comparisons — decoupled fields prevent hand/box clipping (left pair), the dual-agent architecture preserves balance over a ground bar (right pair). (B) Procedurally generated scenes (top) and realistic 3D-FRONT interiors (bottom), all without scene-specific reward tuning (paper Figure 4).

The generalization study scales teachers to 75 scenes and evaluates the distilled generalist on 50 unseen environments. Scene similarity is measured in the method's own representation: SGF and OGF of an unseen scene and a training scene are discretized at $(0.04\,\mathrm{m})^{3}$ resolution, guidance vectors at corresponding positions are compared by cosine similarity averaged across both fields and all positions, and the maximum over training scenes defines the unseen scene's similarity. Success rate declines and mDist grows as scenes become more novel, but gracefully: 80.1% average SR (floor around 70%) and 0.270 m average mDist, with least-squares fit RMSE of 6.1 and 0.092 respectively.

Figure 5: 50 unseen scenes. (A) success rate versus training-set similarity, mean 80.1%, fit RMSE 6.1; (B) mDist versus similarity, mean 0.270 m, fit RMSE 0.092 (paper Figure 5).

Table 2: Generalist statistics on unseen scenes (read from paper Figure 5 annotations and text).

Shape results are almost suspiciously flat: cuboid 88.5 / 0.130, cylinder 86.7 / 0.127, sphere 87.2 / 0.131 — within 1.8 points on SR and identical to three decimals on mDist. That flatness is the design paying off: shape enters the policy only through where the eight keypoints sit, while observation layout, network, and action space remain untouched and only two minor reward terms adapt.

Table 3: Shape-specific policies on the same 20 scenes; the cuboid row equals our method in Table 1 (paper Table II).

Figure 6: Cylinder (left) and sphere (right) transportation in procedural and realistic scenes with only minimal shape-specific reward tuning (paper Figure 6).

On the real Unitree G1, each method gets 6 trials per obstacle type; a shared pickup policy lifts the box first, then the transportation policy takes over, both at 50 Hz. HOTICE scores 6/6 on side, 5/6 on overhead, and 5/6 on ground obstacles — 16 of 18 — against 12/18 for the best baseline. Two details deserve attention. The ranking of components flips relative to simulation: Single-Agent & HOD-PF (11/18) loses to Dual-Agent & SGF (12/18), suggesting that on real hardware stability (the dual-agent contribution) is scarcer than field decoupling. And ground obstacles are hardest for everyone (worst case 1/6), because hurdling demands extreme single-leg support postures. Figure 7 makes all three stories visible: HOTICE threads a side pole, an office chair, and a ground tube; Dual-Agent & SGF collides with the side box; Single-Agent & HOD-PF catches the ground tube and falls.

Table 4: Real-world results, 6 trials per obstacle type (paper Table III). With 6 trials per cell, one trial is worth 16.7 percentage points, so single-trial gaps are not statistically significant and the paper reports no confidence intervals.

Figure 7: Real rollouts. Top: HOTICE clearing a side pole, an office chair, side boxes, and a ground tube. Middle: Dual-Agent & SGF colliding with the side obstacle. Bottom: Single-Agent & HOD-PF striking the ground obstacle and falling (paper Figure 7).

Limitations

The authors list four. Post-collision behavior is not modeled: a disturbance beyond the robot's recovery capacity ends in a fall. Richer carrying strategies — shouldering the object, tucking it against the hip — remain out of reach. Dynamic scenes with moving obstacles are untouched. And "acceptable" collisions plus active obstacle interaction — leaning on a wall to thread a tight gap, brushing a bean bag without penalty, kicking small obstacles aside with the feet, pushing a door with the back — are left as future work.

Three further judgments of my own. First, the perception loop is not closed: scenes must be pre-mapped by SLAM and the box must wear AprilTags for pose estimation, making this a localization-dependent system rather than a perception-complete one; the authors themselves note that live LiDAR reconstruction is blocked by the single-LiDAR field of view. Second, the generalization metric is circular by construction: unseen-scene similarity is defined in the method's own SGF/OGF representation, so "looks similar" and "encodable by the representation" are nearly the same statement; an external similarity measure would test the claim more honestly. Third, payload and grasp scope are narrow: 3.5 kg and a 40 cm cube cap with the object assumed already held leaves grasp synthesis, heavy loads, and off-center payloads out of scope, while field hyperparameters $d_{0}$, $\eta$, $\xi$ are inherited and the 8 cm margin is justified only by a balance argument, with no sweep reported.

Conclusion and Outlook

HOTICE answers "how should a humanoid carry things through clutter" in three layers: encode geometry as two decoupled potential fields so robot and object each get a map; split control into dual agents with separate reward streams kept coupled by shared observations and shared rewards; and consolidate experience by distilling 75 scene specialists into one deployable generalist. The numbers back each layer: 88.5% / 0.130 m over 20,000 simulated trials, 80.1% / 0.270 m on 50 unseen scenes, and 16 of 18 real trials.

The open directions are equally legible: post-collision recovery, dynamic obstacles, acceptable contact and active obstacle interaction, richer carrying postures, and a perception loop free of AprilTags and pre-mapping. For humanoids heading into warehouses and homes, "carrying something through a doorway" is a more frequent daily act than empty-handed parkour — HOTICE turns that daily act into a trainable, distillable, deployable control problem for the first time.

Golden Line

The robot may lift its feet only at the last instant before a bar, but the box in its arms must start yielding half a meter early — HOTICE writes that time difference into two potential fields, and lets the hands claim credit only for the residual adjustment that is truly theirs.