Research map / 3D vision / 3D generation
LessMimic: Versatile Humanoid-Object Interaction with Unified Distance Field Representations
Yutang Lin, Jieming Cui, Yixuan Li, Wei Liang, Baoxiong Jia, Yixin Zhu, Siyuan Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LessMimic, a humanoid-object interaction approach using unified distance field representations to enable versatile whole-body interaction without motion references or task-specific rewards.
Problem
Existing humanoid-object interaction methods depend on motion-reference inputs or task-specific rewards, which couple policies to particular motion scripts, object geometries, and contact timings, limiting geometric generalization and flexible interaction control.
Approach
Uses unified distance field representations for humanoid-object interaction so that simple commands can specify motion intent, removing reliance on motion references and task-specific rewards.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AIP interaction reward (ablation) | LessMimic · Carry and sequential N=5 | success rate (%): 0.0 -> 82.9 (Carry); 5.2 -> 62.1 (N=5) | Improved |
| add AIP interaction reward (ablation) | LessMimic · Push contact / SitStand | contact rate / success: Push 1.0x: 72.5 -> 51.3; SitStand 1.0x: 98.3 -> 80.3 | Hurt |
| add synthetic command augmentation (ablation) | LessMimic · Sequential multi-skill, N=5 | success rate (%): 22.1 -> 62.1 | Improved |
| add geometry randomization in RL post-training (ablation) | LessMimic · PickUp, scale 1.0x | success rate (%): 0.0 -> 100.0 | Improved |
| add RL post-training after behavior cloning (ablation) | LessMimic · PickUp, scale 1.0x | success rate (%): 31.7 -> 100.0 | Improved |
| use Transformer instead of MLP backbone (ablation) | LessMimic · Sequential multi-skill, N=5 | success rate (%): 1.7 -> 62.1 | Improved |
| encode DF history with VAE latent instead of raw DF (ablation) | LessMimic · PickUp overall | success rate (%): 81.2 -> 89.0 | Improved |
| use DF instead of bounding-box or voxel observations (ablation) | LessMimic · PickUp overall | success rate (%): 52.5 / 66.5 -> 81.2 (raw DF) | Improved |
| use DF-based policy vs motion-conditioned baselines at out-of-range scale 0.4x | LessMimic · PickUp, scale 0.4x | success rate (%): 0.0 -> 63.0 | Improved |
| motion-conditioned baselines vs LessMimic at nominal scale | HDMI/ResMimic · SitStand and Push, scale 1.0x | success / contact rate: ResMimic SitStand 100.0; HDMI Push 97.3 vs ours 80.3 / 51.3 | Mixed |
| replace MoCap DF input with egocentric depth (vision distillation) | LessMimic · Sequential multi-skill, N=5 | success rate (%): 62.1 -> 15.9 | Hurt |
| add rear depth camera to vision policy (ablation) | LessMimic (Vision) · SitStand | success rate (%): 0 -> 31 overall | Improved |
Closest papers by meaning
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.89)
- HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning (NeurIPS 2026 · similarity 0.87)
- Omnigrasp: Grasping Diverse Objects with Simulated Humanoids (NeurIPS 2024 · similarity 0.87)
- Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.87)
- CooHOI: Learning Cooperative Human-Object Interaction with Manipulated Object Dynamics (NeurIPS 2024 · similarity 0.87)
- SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
