BRIC: Bridging Kinematic Plans and Physical Control at Test Time

TL;DR

BRIC addresses long-horizon human motion generation by bridging diffusion-based kinematic planning and physics-based control through test-time adaptation and signal-space guidance. It introduces two key innovations: an online forgetting-aware adaptation of the motion policy (L_CF) plus a robustness objective (L_Robust) to align the controller with noisy planner outputs, and a lightweight, backprop-free test-time guidance that optimizes task objectives directly in the signal space. By jointly applying adaptation and guidance, BRIC achieves physically plausible, coherent long-term motion across text-driven, goal-reaching, obstacle avoidance, and human-scene interaction tasks, outperforming prior approaches. The method offers practical benefits for robotics and animation by enabling robust long-term control under distribution shifts without expensive retraining or diffusion-model backpropagation.

Abstract

We propose BRIC, a novel test-time adaptation (TTA) framework that enables long-term human motion generation by resolving execution discrepancies between diffusion-based kinematic motion planners and reinforcement learning-based physics controllers. While diffusion models can generate diverse and expressive motions conditioned on text and scene context, they often produce physically implausible outputs, leading to execution drift during simulation. To address this, BRIC dynamically adapts the physics controller to noisy motion plans at test time, while preserving pre-trained skills via a loss function that mitigates catastrophic forgetting. In addition, BRIC introduces a lightweight test-time guidance mechanism that steers the diffusion model in the signal space without updating its parameters. By combining both adaptation strategies, BRIC ensures consistent and physically plausible long-term executions across diverse environments in an effective and efficient manner. We validate the effectiveness of BRIC on a variety of long-term tasks, including motion composition, obstacle avoidance, and human-scene interaction, achieving state-of-the-art performance across all tasks.

Figures (12)

• Figure 1: BRIC introduces a test-time adaptation (TTA) framework for physically plausible long-term motion generation, effectively and efficiently bridging the execution gap between autoregressive diffusion-based kinematic planners and physics-based controllers. It enables robust execution of highly extended and challenging tasks, such as (left) reliable obstacle avoidance over 100$\,\mathrm{m}$ and (right) navigation through an indoor environment spanning about 12 minutes for human-scene interaction, with the visit order of rooms annotated in the figure. Animated videos of these tasks, along with additional visualizations, are available on the project page at https://bric2026.github.io/.
• Figure 2: The overall procedure. (Left) Test-time adaptation and guidance bridge the distribution gap between the motion planner and physics based controller in an autoregressive manner. (Center) Adaptation mitigates catastrophic forgetting ($\mathcal{L}_{\text{CF}}$) and improves robustness ($\mathcal{L}_{\text{Robust}}$), while guidance refines motion plans by optimizing task objectives. (Right) Before adaptation, the agent fails to track motion plans indicated by red dots. After adaptation, it performs robust and successful motions.
• Figure 3: Comparisons on goal-reaching. (Left) Standard setting and (Right) long-term setting, each evaluated over two target distance ranges. MaskedMimic$_{\{\mathrm{R}, \mathrm{P}\}}$ report results for the "reach" and "path-following" modes, respectively, as described in tessler2024maskedmimic.
• Figure 4: Comparisons on the HSI task. The horizontal axis indicates the sequence of rooms visited by the agent according to the scene plan. Models marked with "${\#}$" are finetuned for the HSI task xiao2024unifiedtevet2025closd.
• Figure 5: Qualitative comparison between CLoSD (red) and BRIC (blue). (Top) In the obstacle avoidance task, CLoSD fails to navigate around obstacles, while our method succeeds under the same conditions. Motion direction is indicated by yellow arrows. (Bottom) In the HSI task, CLoSD becomes stuck and fails to sit, whereas BRIC successfully completes SIT, GETUP, and REACH actions (from right to left).