Skip to content

Back to the catalog

From Digital Human Interactions to Physics-Based Humanoid Skills: Physics-Grounded Post-Training of Interaction Generators
arxiv.orgpaper

From Digital Human Interactions to Physics-Based Humanoid Skills: Physics-Grounded Post-Training of Interaction Generators

The paper introduces DIGHT, a co-adaptive framework that couples a text-conditioned digital-human interaction generator with a physics-based humanoid tracking policy to produce physically plausible two-humanoid interactions. Rather than collapsing executability and interaction-fidelity signals into one scalar reward, it applies physics-decoupled diffusion Direct Preference Optimization (DPO) to align the generator using criterion-specific preferences derived from tracking error, friction, floating, and simulator force feedback (contact occurrence, location, duration, magnitude), then uses the aligned generator's outputs to fine-tune the tracker. Experiments show improved physical plausibility and more reliable, faithful simulated humanoid interactions compared to prior tracking-based approaches.

TL;DR DIGHT aligns a text-to-motion humanoid interaction generator with physics-grounded preferences via diffusion DPO, improving both executability and contact fidelity in simulated two-humanoid interactions.
Open original ↗
Document

What it claims

DIGHT post-trains a text-to-two-person-motion diffusion generator using feedback from a physics-simulated tracking policy, then fine-tunes that tracker on the aligned generator's outputs. It was tested with two generator backbones (InterGen, TIMotion) on two datasets (InterHuman, Inter-X). Across all four settings, the authors say it lowers all five evaluated physical errors while "largely preserving" motion quality and text alignment. On a new contact-annotated subset of the InterHuman test set, it raises contact-topology precision from 38.29% to 46.77% and recall from 52.16% to 63.94% over the InterGen baseline. The authors also report that it generalizes to temporally compositional and multi-role interactions.

How it works

  • Tracker. A PHC-style RL policy drives a full SMPL humanoid through PD controllers. It takes the current state and the next reference state and outputs target joint rotations. On top of PHC's tracking rewards it adds an interaction-graph reward that preserves the spatial relations between the two agents.
  • Retargeting. Motions are retargeted to one shared simulation morphology. A geometry-based contact refinement then fixes contacts that drift during retargeting, such as a palm contact landing on the wrist or forearm.
  • Sampling. For each training prompt, the generator samples K candidates. The frozen tracker executes each candidate, and it also executes the ground-truth motion. The ground-truth rollout supplies the reference contact pattern, but it is not automatically used as the preferred sample.
  • General preferences. Three preference sets cover tracking deviation, frictional dissipation and floating. To win a pair, a candidate needs lower error on that criterion and better motion-text alignment.
  • Contact preferences.
  • Simulator contact forces between the two humanoids are grouped by pairs of body regions. There are six regions: core, head, arms, hands, legs and feet.
  • A sliding window computes the share of frames in which each region pair is in contact. A region pair counts as a persistent contact if that ratio reaches a threshold in at least one window. This filters out one-off collisions but tolerates brief gaps.
  • The resulting set of persistent contacts is compared with the ground-truth rollout's set to build preferences on contact occurrence and contact quality (location, duration, force magnitude).
  • Generator alignment. The five preference sets drive a "physics-decoupled" diffusion DPO. Each criterion keeps its own supervision instead of being merged into one scalar reward, and no gradients pass through the simulator. Both backbones use clean-motion (x₀) prediction.
  • Tracker adaptation. The aligned generator is frozen, and its samples become reference motions for fine-tuning the tracker on its original objective.
flowchart LR
  G[Generator] -->|K candidates| T[Frozen tracker in sim]
  GT[Ground-truth motion] --> T
  T --> P1[Tracking / friction / float prefs]
  T --> P2[Contact occurrence + quality prefs]
  P1 --> D[Decoupled diffusion DPO]
  P2 --> D
  D --> G2[Aligned generator]
  G2 -->|reference motions| FT[Fine-tune tracker]

Evidence

  • Contact benchmark (InterHuman test subset, labelled with whether contact is required and which body-region pairs should touch), DIGHT vs InterGen:
  • precision 38.29% → 46.77%
  • recall 52.16% → 63.94%
  • Physical errors: the authors claim reductions in all five errors across InterHuman/Inter-X × InterGen/TIMotion. The fetched text gives no per-setting numbers for these.
  • Qualitative comparison: direct tracking can lose the hand-to-hand contact in a handshake. Both direct tracking and InterAgent can produce excessive shoulder-pat forces that destabilize the receiving humanoid.

Limits

  • The fetched text stops partway through the method section. The tables, ablations (including any test of decoupled vs. scalar reward), compute, K and the window thresholds are not visible.
  • "Largely preserving" motion quality suggests some quality or text-alignment trade-off. Its size is not shown here.
  • All results are in simulation on an SMPL humanoid. Nothing indicates real-robot transfer.
  • Contact-topology numbers are reported only on InterHuman against InterGen.
  • The page gives a project page (https://1233chen.github.io/DIGHT/) but states no code release, weights or licence.

Technologies named here

  • diffusion model
  • Direct Preference Optimization

Techniques named here

  • DIGHT
  • Physics-decoupled diffusion Direct Preference Optimization

Related in research