TLDR: A novel framework called Symbolically-Guided Decision Transformer (SGDT) is introduced for deployable multi-robot collaboration. It combines a neuro-symbolic planner, which generates high-level symbolic subgoals using an open-source LLM, with a goal-conditioned decision transformer (GCDT) for low-level robot manipulation. This hierarchical approach improves decision-making, interpretability, and generalizability, demonstrating high success rates and adaptability in complex multi-robot tasks like sandwich making and grocery packing, even with limited or unseen data.
Multi-robot systems (MRSs) are becoming increasingly vital across various high-impact sectors, including construction, manufacturing, surgical procedures, and agriculture. These systems enhance efficiency, robustness, and adaptability by allowing multiple robots to work together in shared environments. However, deploying advanced artificial intelligence, particularly reinforcement learning (RL), in these real-world scenarios faces significant hurdles. RL often demands vast amounts of data and relies on assumptions that limit its effectiveness in complex, long-term tasks like multi-robot manipulation.
Addressing these challenges, a promising alternative known as Decision Transformers (DTs) has emerged. DTs reframe reinforcement learning as a sequence modeling problem, allowing policies to be trained from existing data without needing to estimate value functions. While DTs have shown great potential in single-robot operations, their application to complex multi-robot manipulation has remained largely unexplored until now.
Researchers have introduced a novel framework called the Symbolically-Guided Decision Transformer (SGDT). This innovative approach integrates a neuro-symbolic mechanism with a causal transformer to enable more deployable and effective multi-robot collaboration. The SGDT framework operates hierarchically, combining two main components: a neuro-symbolic planner and a goal-conditioned decision transformer (GCDT).
The neuro-symbolic planner acts as the ‘brain’ for high-level task planning. It generates a sequence of symbolic subgoals, essentially a step-by-step plan for the robots. This planner leverages the Planning Domain Definition Language (PDDL) and an open-source large language model (LLM) like LLaMA3. This design choice makes the system cost-effective and ensures reproducibility, crucial for industrial applications. The planner’s output is structured, interpretable, and can even incorporate human guidance through natural language, allowing operators to supervise and intervene when necessary.
Guided by these high-level symbolic subgoals, the Goal-Conditioned Decision Transformer (GCDT) handles the low-level, sequential decision-making for robot manipulation. The GCDT translates the symbolic subgoals into numerical targets, such as the expected 3D coordinates of the robot arms. It then determines the precise movements and actions required to achieve these targets. By conditioning on these subgoals, the GCDT can manage longer-range temporal dependencies, overcoming a common limitation of traditional RL methods.
The SGDT framework was rigorously evaluated using RoCoBench, a simulation platform designed for multi-robot collaboration tasks. The evaluations focused on two representative tasks: sandwich making and grocery packing. In the sandwich-making task, humanoid and UR5e robots collaborated to prepare sandwiches with varying numbers of items. In the grocery packing task, Panda and UR5e robots worked together to pack items into a box.
The results were highly encouraging. The neuro-symbolic planner demonstrated exceptional reliability, achieving 100% accuracy in sandwich-making tasks and 90% in grocery packing, with minimal need for replanning. This highlights its robustness in generating valid task plans for collaborative settings.
Furthermore, the GCDT’s performance significantly improved when conditioned on high-level subgoals. For instance, in 5-item sandwich-making tasks, the success rate jumped from 43.08% without subgoals to 95.33% with subgoals. This validates the SGDT’s design in strengthening long-term temporal memory and maintaining efficiency even in complex, extended manipulation sequences.
The framework also showed strong capabilities in few-shot cross-task adaptation, meaning it could efficiently adapt to new tasks with limited fine-tuning data. For example, when pretrained on sandwich-making data and fine-tuned with just 100 samples of grocery-packing data, its success rate exceeded 90%. SGDT also demonstrated robust zero-shot generalization, performing well on unseen task configurations without any additional training. For more details, you can read the full research paper here.
Also Read:
- Bridging the Language Gap: How AI is Enabling Robots to Understand Human Instructions
- Self-Guided Action Diffusion: A Leap in Robot Policy Efficiency
This work represents a significant step in applying Decision Transformer technology to multi-robot collaborations. The SGDT framework offers a cost-effective, task-oriented, and highly effective solution for deployable multi-robot systems, with broad potential across various industrial domains. Future work will focus on evaluating and deploying SGDT in real-world industrial settings.
