BAKU: An Efficient Transformer for Multi-Task Policy Learning

Siddhant Haldar, Zhuoran Peng, Lerrel Pinto

Abstract

Training generalist agents capable of solving diverse tasks is challenging, often requiring large datasets of expert demonstrations. This is particularly problematic in robotics, where each data point requires physical execution of actions in the real world. Thus, there is a pressing need for architectures that can effectively leverage the available training data. In this work, we present BAKU, a simple transformer architecture that enables efficient learning of multi-task robot policies. BAKU builds upon recent advancements in offline imitation learning and meticulously combines observation trunks, action chunking, multi-sensory observations, and action heads to substantially improve upon prior work. Our experiments on 129 simulated tasks across LIBERO, Meta-World suite, and the Deepmind Control suite exhibit an overall 18% absolute improvement over RT-1 and MT-ACT, with a 36% improvement on the harder LIBERO benchmark. On 30 real-world manipulation tasks, given an average of just 17 demonstrations per task, BAKU achieves a 91% success rate.

Introduction

The quest for "generalist" robots—agents capable of performing a wide array of tasks—is hindered by the high cost of data acquisition. Unlike domains like natural language processing or computer vision, where data can be scraped from the web, robotics requires physical execution in the real world. This makes every expert demonstration a precious resource.

Current multi-task architectures often face two primary hurdles: "negative transfer," where learning one task degrades performance on another, and low data efficiency, requiring massive datasets to match the performance of single-task experts. BAKU addresses these challenges through a streamlined, transformer-based architecture designed to maximize the utility of limited demonstration data.

The BAKU Architecture

BAKU is designed around a "simple yet meticulous" philosophy, focusing on efficient information fusion and flexible action generation. The architecture consists of three primary components:

1. Multimodal Sensory Encoders

To process diverse inputs, BAKU utilizes specialized encoders for different modalities:

  • Vision: A ResNet-18 backbone processes visual input. Crucially, it employs Feature-wise Linear Modulation (FiLM), allowing text instructions to condition visual features. This ensures the model focuses on task-relevant visual cues.
  • Language: A 6-layer MiniLM encodes natural language instructions.
  • Proprioception: A simple Multi-Layer Perceptron (MLP) processes the robot's internal state.

2. Unified Observation Trunk

Instead of a complex encoder-decoder structure, BAKU employs a causal transformer decoder as its central trunk. Every input—vision, text, and state—is treated as a token. This unified representation allows the model to naturally fuse information across different modalities and time steps, creating a coherent understanding of the environment and the task at hand.

3. Modular Action Heads and Chunking

BAKU separates the "understanding" (observation trunk) from the "acting" (action head). This modularity allows the model to be paired with various action-generation frameworks, including Gaussian Mixture Models (GMM), Behavior Transformers (BeT), and Diffusion Policies.

To ensure smooth and consistent motion, BAKU utilizes action chunking. Rather than predicting a single next action, the model predicts a sequence (or "chunk") of future actions. These predictions are blended using exponential temporal averaging, which mitigates jitter and helps the robot maintain progress toward a goal even if individual observations are noisy.

Experimental Results

Simulation Benchmarks

BAKU was evaluated across 129 simulated tasks, including the challenging LIBERO-90 suite, Meta-World, and DeepMind Control (DMC).

  • LIBERO-90: BAKU achieved a 90% success rate, representing a 36% absolute improvement over established baselines like RT-1 and MT-ACT.
  • Overall Performance: Across all 129 tasks, BAKU demonstrated an average 18% absolute improvement over prior state-of-the-art models.

Real-World Performance

In a real-world kitchen environment involving 30 diverse tasks (e.g., opening ovens, picking up fruit), BAKU achieved a 91% success rate. This was accomplished with an average of only 17 demonstrations per task, highlighting the architecture's extreme data efficiency. The authors found that while simple MLP heads work well in simulation, a VQ-BeT head—which uses vector quantization to handle multimodal action distributions—provided a 5% performance boost in the more complex real-world setting.

Key Insights

  • Model Scaling: Interestingly, larger is not always better in data-constrained robotics. A 114M parameter version of BAKU performed worse than the standard 10M parameter version on LIBERO, likely due to overfitting on the limited demonstration data.
  • Temporal Dynamics: While the transformer can process long histories, the authors found that using only the most recent observation while predicting a long chunk of future actions yielded the best results.
  • Ablation Studies: Removing action chunking led to a 14% drop in success, confirming its necessity for stable manipulation.

Conclusion and Future Work

BAKU demonstrates that a well-designed, relatively small transformer architecture can significantly outperform larger, more complex models in multi-task robot learning. By focusing on efficient multimodal fusion and modular action generation, it sets a new standard for data-efficient imitation learning.

Future work will focus on addressing precision bottlenecks in high-tolerance tasks and exploring "skill chaining" to enable the model to solve long-horizon problems by sequencing the individual skills it has mastered. Additionally, investigating the architecture's ability to generalize to entirely unseen tasks remains a key area for exploration.

Summary updated

Abstract

Training generalist agents capable of solving diverse tasks is challenging, often requiring large datasets of expert demonstrations. This is particularly problematic in robotics, where each data point requires physical execution of actions in the real world. Thus, there is a pressing need for architectures that can effectively leverage the available training data. In this work, we present BAKU, a simple transformer architecture that enables efficient learning of multi-task robot policies. BAKU builds upon recent advancements in offline imitation learning and meticulously combines observation trunks, action chunking, multi-sensory observations, and action heads to substantially improve upon prior work. Our experiments on 129 simulated tasks across LIBERO, Meta-World suite, and the Deepmind Control suite exhibit an overall 18% absolute improvement over RT-1 and MT-ACT, with a 36% improvement on the harder LIBERO benchmark. On 30 real-world manipulation tasks, given an average of just 17 demonstrations per task, BAKU achieves a 91% success rate. Videos of the robot are best viewed at https://baku-robot.github.io/.
BibTeX
@inproceedings{haldar2024baku,
  author = {Siddhant Haldar and Zhuoran Peng and Lerrel Pinto},
  title = {{BAKU: An Efficient Transformer for Multi-Task Policy Learning}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2024},
  url = {https://openreview.net/forum?id=uFXGsiYkkX},
}
Figure