DiffGen: Robot Demonstration Generation via Differentiable Physics Simulation, Differentiable Rendering, and Vision-Language Model

TL;DR

DiffGen addresses scalable robot demonstration data generation by removing reliance on hand crafted rewards and heavy RL training. It builds a differentiable pipeline that fuses a physics simulator $s_{t+1}=f(s_t,a_t)$, a differentiable renderer $I_t=g(s_t)$, and a vision-language encoder $h$, to minimize the distance $L(z^l,z^{I})= - (z^l\cdot z^{I})/(||z^l||\,||z^{I}||)$ between the instruction embedding $z^l=h(l)$ and the observation embedding $z^{I}=h(I)$ via gradient descent on the action sequence $\{a_t\}$. It introduces an episodic long-horizon optimization strategy to mitigate gradient issues and demonstrates zero-shot goal specification and cross-embodiment generalization, validating the approach on three manipulation tasks. The results show higher efficiency and lower human effort than RL baselines, suggesting DiffGen can scale up robot data for future research.

Abstract

Generating robot demonstrations through simulation is widely recognized as an effective way to scale up robot data. Previous work often trained reinforcement learning agents to generate expert policies, but this approach lacks sample efficiency. Recently, a line of work has attempted to generate robot demonstrations via differentiable simulation, which is promising but heavily relies on reward design, a labor-intensive process. In this paper, we propose DiffGen, a novel framework that integrates differentiable physics simulation, differentiable rendering, and a vision-language model to enable automatic and efficient generation of robot demonstrations. Given a simulated robot manipulation scenario and a natural language instruction, DiffGen can generate realistic robot demonstrations by minimizing the distance between the embedding of the language instruction and the embedding of the simulated observation after manipulation. The embeddings are obtained from the vision-language model, and the optimization is achieved by calculating and descending gradients through the differentiable simulation, differentiable rendering, and vision-language model components, thereby accomplishing the specified task. Experiments demonstrate that with DiffGen, we could efficiently and effectively generate robot data with minimal human effort or training time.

Figures (4)

• Figure 1: With DiffGen, we could automatically and efficiently generate robot demonstrations, given simulated task environments and text-based instructions.
• Figure 2: The overall pipeline of our proposed DiffGen. Our system first initiates action sequences, simulates the state changes, and renders visual observations after manipulation, via differentiable simulation and differentiable rendering. Then, a vision-language model is employed to measure the distance between the visual observations and the text-based instructions. Thanks to the differentiability of each component, feasible action sequences can be generated by gradient descent optimization.
• Figure 3: Visualization of the tasks, generated by the PyBullet renderer.
• Figure 4: Loss Curves on Cube-Selection Task