Breadcrumbs to the Goal: Goal-Conditioned Exploration from Human-in-the-Loop Feedback

Exploration and reward specification are fundamental and intertwined challenges for reinforcement learning. Solving sequential decision-making tasks requiring expansive exploration requires either careful design of reward functions or the use of novelty-seeking exploration bonuses. Human supervisors can provide effective guidance in the loop to direct the exploration process, but prior methods to leverage this guidance require constant synchronous high-quality human feedback, which is expensive and impractical to obtain. In this work, we present a technique called Human Guided Exploration (HuGE), which uses low-quality feedback from non-expert users that may be sporadic, asynchronous, and noisy. HuGE guides exploration for reinforcement learning not only in simulation but also in the real world, all without meticulous reward specification. The key concept involves bifurcating human feedback and policy learning: human feedback steers exploration, while self-supervised learning from the exploration data yields unbiased policies. This procedure can leverage noisy, asynchronous human feedback to learn policies with no hand-crafted reward design or exploration bonuses. HuGE is able to learn a variety of challenging multi-stage robotic navigation and manipulation tasks in simulation using crowdsourced feedback from non-expert users. Moreover, this paradigm can be scaled to learning directly on real-world robots, using occasional, asynchronous feedback from human supervisors.

Paper

Similar papers

Peer review

Reviewer 4hMa5/10 · confidence 4/52023-06-14

Summary

The paper introduces Human Guided Exploration (HUGE), a system designed to integrate human feedback within the Goal-Conditioned Reinforcement Learning (GCRL) environment, offering a cost-effective and straightforward method for learning diverse robotic tasks through actual human interaction. The authors' primary objective is to strike a balance between overexploration and underexploration in GCRL tasks by incorporating human-in-the-loop assistance. Consequently, human feedback is utilized to ascertain the states requiring further exploration, building on the traditional GCRL algorithms and the GO Explore exploration framework. In particular, a target distance estimation function is trained using binary feedback, facilitating the generation of a more precise subgoal. Subsequently, the Goal-Conditioned Supervised Learning (GCSL) training paradigm is implemented to establish an interaction strategy with the environment. Experimental results in Mujoco and Pybullet demonstrate that HUGE outperforms the preceding GCRL algorithm and Human-in-the-loop algorithm.

Strengths

- The integration of human feedback into the GCRL setting is well-conceived. Hence, the motivation is clear, and the proposed method is, in general, logical and sensible. - This technique facilitates a simple interface between the human labeler and the algorithm, in which the human supervisor provides binary evaluations to establish which state-goal pairings are closer in comparison to others.

Weaknesses

- The paper presents a degree of innovation that is somewhat restrained, as it essentially combines modified versions of existing methodologies (GCRL, Go Explore, and Human Preference). The foundational GCRL algorithm used is the previously established GCSL, while the exploration component is derived from the Go-Explore paradigm. It is important to note that GCSL paired with Go-Explore could have operated independently without the need for human feedback. - As per the previous discussion, it's not unexpected to see improved results when human feedback is integrated into the Go-Explore process, as it introduces more a priori knowledge. The implementation of binary feedback can help to decrease the exploration space by learning to rank and establishing a subgoal model. - This strategy essentially transforms the behavior cloning challenge into a binary human feedback problem. While this appears logical, it makes the motivation somewhat ambiguous and makes specific presumptions about the task at hand. Consequently, the work lacks a degree of originality and novelty. The three fundamental components of the proposed method: the "goal-conditioned policy", "hindsight relabeling", and "learning reward models" from human preference, have all been previously suggested in earlier papers. - I think the tasks for evaluation are borderline novel and perhaps not hard enough when we consider a setting where human feedback can be utilized (these tasks are still challenging for RL-based methods learning with dense/sparse rewards). The maze tasks, robotic manipulation tasks are commonly seen in literature (in fact I am surprised PPO is completely failing on some of these tasks, I have personally trained PickNPlace tasks with a RL-based method with a properly shaped reward - Otherwise the authors can also explore other SOTs such as TD3 or SAC) . These tasks are not seen as a fundamental challenge for present learning methods. The main result of the paper is comparison of the proposed method against RL algorithms like PPO and LEXA in terms of sample efficiency. I prefer to see longer horizon tasks where an RL agent with a dense reward cannot easily solve. I also did not see comparison results for the Sim2Real tasks nor for PickNPlace.

Questions

Other comments: - The spelling sparce to sparse in Figure 5.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

- The work lacks novelty.

Reviewer erJy7/10 · confidence 4/52023-07-03

Summary

This paper focuses on the exploration problem in decision-making tasks. Previous works try to leverage human guidance with constant synchronous high-quality human feedback, which is expensive and impractical to obtain. In this paper, the authors propose Human Guided Exploration (HUGE), which is able to leverage low-quality feedback from non-expert users. Specifically, the key idea is to separate the challenges of directed exploration and policy learning. Human feedback is only used to direct exploration, while self-supervised policy learning is used to independently learn unbiased behaviors from the collected data. The task for human annotators is just to select which of two states is closer to a particular goal. Then this ranking will be used to train a distance function for goal selection. The experimental results on robotic navigation and manipulation tasks in both simulation and real-world robots demonstrate the advantages of the proposed method.

Strengths

1. The idea of the proposed method is simple yet effective. Separating the exploration and policy learning dramatically reduces the annotation effort and influence of noisy annotation on policy learning. Using human preference to learn the distance function is also better than directly learning the reward function. 2. Thorough evaluation of multiple environments, including both simulation and real-world tasks in robotic navigation and manipulation domains. The experimental results show that HUGE outperforms other methods by a large margin.

Weaknesses

1. Figure 1 provides limited information. It is hard to find that human feedback is noisy and asynchronous. The tasks of pick-and-place and drawing do not contain detailed explanations. This figure can be improved by adding a comparison between previous works that require high-quality data and their method. 2. The method may need to be evaluated on more complex real-world tasks. The authors argue that Novelty-based exploration performs task-agnostic exploration of the entire state-space, thereby over-exploring the environment. But it seems that the environments used in this paper do not contain very large exploration space.

Questions

No particular questions.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No limitation is discussed in the paper. One potential limitation is that the human annotation of the preference of states may be difficult for tasks that require logic reasoning. The experiment environments used in this paper mainly focus on L2 distance, which is easy for humans to quickly select the state close to the goal.

Reviewer EF3C7/10 · confidence 3/52023-07-06

Summary

This paper introduces HUGE, a Reinforcement Learning algorithm that makes use of human preferences to guide the selection of partial goals. HUGE expands upon Goal-Conditional Supervised Learning (GCSL) by improving the goal selection method using noisy labels from humans to form a model, $f_\theta$, of distances to a goal state. This model $f_\theta$ is then used set goals for the GCSL algorithm to learn in a self-supervised fashion, thus bootstrapping learning. The paper features an extensive evaluation of HUGE, with 6 synthetic tasks, 2 real robot tasks and one task trained from crowd-sourced feedback. Results show that HUGE outperforms a large number of baselines.

Strengths

* HUGE combines intuitions from GCSL and GoExplore, and can take advantage of asynchronous human preferences. * The paper is really well written and is easy to follow. * The paper presents a thorough evaluation. I find it particularly that HUGE Was able to train a real robot policy in under 30 hours with just 130 annotations. * HUGE can take advantage, but does not require, expert demonstrations through Behaviour Cloning.

Weaknesses

* It would have been more representative to choose PEBBLE [1] rather than Christiano et al. (2017) as a more representative example of the performance of Preference-based Reinforcement Learning. * Some of the results are buried in the appendix. For instance section 5.2 does not mention what was the task analysed are. From Figure 6. left, I assume it's the Kitchen task from D4RL. Still, neither the final performance nor the task are mentioned in the main text. * Similarly, the main results of the analysis comparing the final performance and the number of annotations should be included in the main text, since it provides readers with an estimation of how many labels they may need for their task. **Post-rebuttal update** Authors added a comparison against PEBBLE, where HUGE outperforms it.

Questions

1. How were the annotations for the real-world tasks obtained? 2. Why does BC+Ours finish early in Figure 5 Kitchen & Bandu? **Minor nitpicks and suggestions:** * $\mathcal{G}$ is undefined in equation 4. * In algorithm 1, step 5, the second argument to PolicyExploration should be $f_\theta$. * In figure 6 left, what is the difference between "human + 5 demos" and "crowd-source + 5 demos". Moreover, the colours of their curves are indistinguishable to my eyes. * When defining $f_\theta$, I would clarify that "closer" is not necessarily meant as a mathematical distance, but rather as a human intuition of closeness, which I belive is what $f_\theta$ is supposed to capture. **Post-rebuttal update** See discussion for answers to the above questions.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitations have been adequately addressed.

Reviewer MJKZ6/10 · confidence 4/52023-07-07

Summary

Broadly, the paper addresses the challenge of learning multi-stage robotic navigation and manipulation tasks in simulation. The paper frames several benchmark tasks as goal-conditioned RL (GCRL), and they present a novel method ("HUGE") for leveraging human preferences collected during learning. In particular, the preferences are used to train a goal-selector model that guides (biases) exploration. For policy-learning, they use goal-conditioned supervised learning on the collected replay buffer, without requiring a handcrafted task-specific reward. Compared to [1], HUGE also uses hindsight relabeling to learn from a replay buffer despite a sparse reward, but it differs in that it biases sampling from the replay buffer based on human preferences. Compared to [11], HUGE also collects human preferences iteratively (intertwined with learning). However, HUGE doesn't use its collected human feedback to learn a reward or otherwise directly bias the policy. Instead, the feedback is only used to bias exploration. Compared to [17], HUGE also has an exploration phase that builds an archive of trajectories. However, HUGE chooses promising goal states and generates trajectories using its goal-condition policy, while [17] chooses promising start states based on novelty and can generate trajectories with various exploration policies (random, epsilon-greedy, novelty-based). The authors evaluate on four simulated manipulation tasks, two simulated navigation tasks, and two real-robot tasks. They compare to several baselines including the related work mentioned above as well as PPO (with dense and sparse reward variants).

Strengths

Their method doesn't require a handcrafted reward and it's able to leverage noisy, infrequent human feedback. These strengths are relevant and valuable to this domain. Evaluation on a variety of simulated and real tasks. Impressive crowdsourcing infrastructure and diversity of annotators. Interesting, well-explained adaption of real-world tasks to fit their use of goal-conditioned RL.

Weaknesses

Lack of clear, fair numeric results on task performance across HUGE and the baselines: = In Figure 5, I can only compare task performance after a given number of steps. A table with a performance number after training each approach to convergence would be more clear and would better support the claim from your text "HUGE beats prior work on the proposed long horizon tasks in simulation". = It's not clear why some tasks were trained for millions of steps and others not. = The paper text conveys that Figure 5's various approaches are using a poorly-tuned reward function ("we did not do any reward finetuning"). In particular, this makes the PPO results questionable. = This section also uses synthetic human preferences derived from your reward function, instead of real human preferences. This confuses me and contradicts the earlier premise of HUGE. HUGE with this synthetic source of preferences would appear to be another RL method that utilizes a handcrafted reward function and not human feedback. This muddies your comparison with other approaches, especially [11] which is intended to be guided by real (not synthetic) human feedback. = It seems like other baselines might benefit from the addition of behavior-cloning pretraining, as you did for "BC + Ours".

Questions

Can you improve your comparison to baselines based on the feedback I gave above? In particular, I am most interested in a robust, fair comparison to [11] ("Human Preferences"), because I consider that work the most directly comparable to yours: it doesn't require a handcrafted reward signal but it does require collection of human preferences intertwined with learning. I hope you'll address the weaknesses in your PPO baseline I mentioned earlier: train to convergence with a fine-tuned reward function. Figure 5 currently shows that HUGE (with synthetic preferences derived from a reward function) outperforms PPO using the same reward function. If this result holds after you improve the PPO baseline, can you analyze this? Would it mean the synthetic variant of HUGE is fundamentally a better optimization algorithm compared to PPO? I notice your two real-world tasks are fully-observable. How would you apply your approach to a real-world task that is only partially-observable, such as room-scale manipulation with visual perception?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

No concerns here

Reviewer EF3C2023-08-15

Thank you for the rebuttal.

Thank you for your detailed rebuttal, the additional experiments regarding PEBBLE and for really trying to get the PEBBLE baseline working. It makes sense that insufficient unsupervised exploration would be cause PEBBLE to collapse. Since my main concern and questions have all been addressed, I will increase my rating.

Reviewer MJKZ2023-08-16

Thanks for the detailed rebuttal. The main novel contribution here is the use of a noisy preference signal to bias exploration, while avoiding biasing the learned policy. The rebuttal experiments and baselines show more clearly that this works well, regardless of whether this is synthetic (e.g. a poorly-tuned reward function) or real human preferences. I've revised my rating.

Reviewer 4hMa2023-08-20

Response to the Authors

I would like to thank the authors for their response and efforts towards improving the paper. In the light of reading other reviewers opinion and the additional results, I'm inceasing my score to 5.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC