Author | Cen Feng
Under the wave of generative AI, the robot community is undergoing a major transition from traditional control to end-to-end learning. Like many fields, there is also a belief in the scale of data—believing that “the more data, the smarter” is the only path to universal embodied intelligence. However, at the just-concluded IEEE/RSJ International Conference on Intelligent Machines and Systems (IROS), Jen Jen Chung from the University of Queensland presented a thought-provoking conclusion based on her over a decade of research cases in her keynote speech: Not all data is equal; blindly accumulating data can even lead to system crashes and strategy failure. Professor Chung directly addressed the core point at the beginning of her speech: Massive data does not equal high information content. There are many examples of having massive data yet achieving nothing, because many pieces of data simply cannot help us understand the essence of the problem. On the way toward universal autonomous systems, what determines the upper limit of intelligence is not the absolute scale of data, but the quality and information density of the data. At the beginning of her research career with the drone autonomous gliding project, Professor Chung deeply understood that “exploration has a cost.” To make fixed-wing drones use rising air currents independently like birds, they must consume precious energy to explore unknown states and motion spaces. This forced her to think: When to collect data? How to collect data? Blind and uniform sampling not only wastes resources but may also slow down the learning process. These early studies made her realize that data value is not uniformly distributed; collecting high-value information as needed is far better than massive, indiscriminate redundant observations. This deep understanding of the non-uniformity of data value is particularly crucial in distributed multi-robot systems. When the research subject expands from a single unit to multi-agent collaboration, and faced with limited, decaying, and highly delayed real communication channels, “broadcasting all data to all nodes at all times” is not only unrealistic but may even cause catastrophic network congestion. In this scenario, Professor Chung proposed the only golden standard for measuring the value of data sharing: “Can knowing this piece of information lead to a change in behavioral decision-making?” Based on this, her team introduced a highly creative core logic: Transmit the most critical information, not all the data. During the speech, Professor Chung presented a series of cases that perfectly illustrate how to achieve a delicate balance between bandwidth usage and data fidelity. At the end of the speech, Professor Chung pointed out the current hottest end-to-end strategy learning paradigm in robotics, facing the “big elephant in the room.” In complex contact-rich assembly tasks, even when massive multimodal data (visual, force, proprioception) is fed to the robotic arm, its performance is extremely mediocre, and it even mechanically replicates insertion actions in empty scenarios. For this reason, Professor Chung’s team proposed a “strategy repair” solution. Experiments show that under the same teaching data budget, this method of selectively selecting “repair” samples significantly improves strategy generalization performance more than blindly expanding data. This once again confirms the theme of this speech: “Data quality determines the limit; blindly pursuing data scale is just a futile digital game.” Below is the speech by Jen Jen Chung, edited according to the original English speech content by AI Technology Review:
▎Speech Title: (Geometry Strikes Back: Scaling 3D Reconstruction and Structuring 3D Scene Generation)
Speaker: Jen Jen Chung, Queensland University01
Starting with drone gliding: The disconnect between data and information
“Data” is not equivalent to “information”. Some data fragments can carry a high amount of information; however, you might also have a massive amount of data yet still achieve nothing. It is particularly meaningful to re-examine the relationship between “data” and “information” today. In the current technological wave, we are extremely obsessed with collecting as much data as possible, believing that “the more data, the better we can solve all problems”. What I want to do today is to challenge this mindset. The disconnect between data and information has been a theme in many of my previous studies. My doctoral research focused on autonomous gliding for drones. Nature has provided us with an excellent example: birds know how to extract energy from wind fields and use the rising thermal air currents to increase their altitude. What we were exploring at that time was whether fixed-wing drones could perform the same operation. From the perspective of exploration strategies, this is a very interesting topic: the longer the flight time, the more opportunities there are to explore the “state-action space”. However, the key lies in:Exploration requires a cost. You must consume the energy that was painstakingly obtained to explore the unknown areas in the state-action space; and there is no mechanism to ensure that your efforts will pay off. You may spend a lot of effort exploring a specific area, but end up with nothing. Therefore, if you want to learn quickly, you must carefully balance “when to collect data” and “how to collect data”. As long as the scale is appropriate, with just a few training rounds, you can develop a fairly efficient gliding strategy. In the 20th to 27th rounds, we have developed a very good control strategy that enables the drone to obtain energy from the thermal airflow. What is even more interesting is analyzing its value function space. In these charts, each curve shows the value function at different airspeeds. You will find that when the airspeed reaches a certain critical point, the net energy gain drops completely to zero—because the flight speed is too high, and the drone will directly hit the thermal airflow, with no time to capture energy from it. Once the control strategy understands this, it realizes that exploring that area is meaningless. Even if you continue to collect more data in that area, it will not improve the strategy by even a little bit. Another example is the shortest path planning in an unknown cost field. The key issue is finding the optimal path between two points, while the real cost field shown in the upper left corner is completely unknown at the initial state. You must use local sampling to understand the costs of traversing each region. Two points need to be kept in mind at this time: first, sampling has a cost; you must pay a physical price to explore and collect data; second, your ultimate goal is to find the shortest path, rather than precisely reconstructing the entire cost field. Once this is clear, we should not perform uniform sampling, or sample blindly just to reduce the estimation error of the cost field. How can we guide data collection to most efficiently improve the accuracy of inferring the shortest path? Since the starting point and ending point are known, we can first perform a bidirectional heuristic search. We make the following assumption: if we collect observational data at some point right now, how will this measurement change our posterior confidence? How will it reshape our search boundaries? What kind of prior changes will it bring in terms of the measurement type or numerical scale? What is the order of magnitude of the expected value? Most importantly: will it change your judgment on the shortest path direction? If you predict that this measurement cannot change the path estimate at all, then this sampling is completely unnecessary. Accordingly, we can targeted guide data collection. Please see the rapid progression exploration method we proposed in 2017: once the algorithm determines that a certain region is a highly costly area, it will immediately stop sampling the state space and the exploration sub-space—because the probability of finding a shorter path by moving upward-left is extremely low. We should shift our valuable data collection budget to the lower-right corner, where there is a higher potential to discover better solutions. In these two cases—whether it’s finding the shortest path or the autonomous gliding of a drone—collecting redundant data doesn’t provide any real help in solving the problem, but at least it doesn’t ruin the existing solution; it simply slows down the computation efficiency.02
Multi-robot system:
Data sharing has never been free of cost
In certain fields, blindly inserting more data into the system is not only ineffective, but can even cause a rapid deterioration in overall performance. This phenomenon is particularly prevalent in distributed multi-robot systems. The reason is that data distribution within the system entirely depends on the underlying communication infrastructure.Data sharing has never been free of cost“.In reality, you can never avoid the bandwidth bottleneck; communication channels are often filled with limitations, attenuation, and delay. Given the physical constraints of network channels, “broadcasting all data to all nodes at all times and from anywhere” is simply impractical. Additionally, the marginal utility of data is not evenly distributed; it depends heavily on the properties of the data itself, the identity of the receiving node, and the timeliness of reception. Considering these three factors, how should we plan the data distribution mechanism between systems so that each node can obtain the most critical information?” We can start with “communication targets” and “communication timing”. The most intuitive implementation scenario is multi-robot collision avoidance. At what moment do drones need to know the course of other robots? Imagine a swarm of drones with intertwined routes: do they need to maintain communication at all times? Or is communication only necessary when there is insufficient reliable information about the behavior of other individuals? The core question you must ask is:With this piece of information, can it lead to a change in behavioral decisions?“When the nodes turn yellow, it indicates that they are communicating or requesting the other party’s status; when they turn blue, it means the onboard prediction is sufficiently reliable and no additional data is needed. Through this carefully selected communication mechanism, we significantly reduce the network channel load and greatly increase the probability of high-value information reaching its destination accurately. Another aspect of the problem is ‘what is being transmitted’. In addition to directly sending raw data, we can completely use approximate representations.” We have introduced a very clever mathematical tool—the Bloom filter. Imagine a collaborative graph-building exploration task, where the robot needs to synchronize the areas it has traversed and observed on the grid map. While you can directly transmit detailed position trajectories, there is a much more efficient way of encoding them. For example, the Bloom filter uses a hash mapping to compress state information into high-density binary arrays, thus achieving an amazing data compression ratio. Although there is a certain loss in data precision, the trade-off is a very small data volume, which significantly improves the successful transmission rate over harsh channels. When the information is sent to the red robot, it receives the large number of packets and can perform local status queries to verify whether a specific grid cell has been explored. The red robot can determine that the grid (1,7)(1,7) has been observed by the green robot; when querying another grid, if the hash bit is 00, it indicates that the other party has not yet touched it, and that area is then determined as a potential candidate worth exploring. However, the cost of this lossy compression, as I mentioned, is that some accuracy is sacrificed. False positives may occur when querying certain grids—a grid that has not been explored yet shows up in the hash query as if it had been included in the set. This is an inevitable cost of high compression ratios. However, a thorough understanding of the internal mechanism of the tool enables us to skillfully introduce other engineering techniques, thereby breaking the performance limits of the original tool. A minimal yet sophisticated solution is to "salt" the data: that is, introduce a random salt value, combine it with the data, then perform hashing, and broadcast the salt value along with the data. In this way, during continuous observation and communication cycles, we can use distinct and independent salt values each time, and the hash mapping output also varies. This means that the data transmitted in each transmission in the network is statistically independent, representing independently observed samples. This statistical independence allows us to utilize all the theoretical advantages of Bayesian inference and sequential information superposition. This enables the receiving end to make more robust decisions. Even if the system’s preset maximum single false positive rate is as high as 50%, the instantaneous false positive rate may be very concerning, such as the false alarms shown by the white noise points in the figure. However, through cross-period information disentanglement, the posterior confidence error of each robot is significantly reduced, far exceeding the instantaneous false positive level. This allows us to enjoy the channel throughput benefits brought by high data compression ratio, while also maintaining high-quality decision-making capabilities. Therefore, mechanisms such as the Bloom filter provide us with a lever for direct trade-off between communication bandwidth overhead and data accuracy. In multi-machine collaborative exploration tasks, overall efficiency is significantly optimized. I prefer to view it as a quantitative trade-off between “data precision” and “timeliness”—which precisely corresponds to “transmission content” and “transmission timing” in the system’s data flow. But the above solution still has an underlying assumption: the system must receive the entire data packet completely before it can start decoding and using it. This is another bottleneck we are trying to overcome. Currently, the robot communication architecture tends to have a conventional mindset: to utilize any information, one must wait for the complete data packet to be received before unpacking and parsing it. Once faced with a poor or intermittent network environment, as long as there is partial packet loss, the entire transmission will be completely invalid. Instead of getting stuck in a black-and-white transmission dilemma, can we further trade data resolution for time? Can we introduce a streaming transmission mechanism so that we don’t have to wait for the entire data set, or receive lightweight packets first and decode them immediately, thereby enabling timely decision-making? In this way, even if the blue robot at the edge of weak signal only receives half of the packets, some of the information it gets can still be directly used for collaborative planning. If you experienced the internet in the 90s like I did, you may still have a vivid memory of the streaming media mechanism of that time. We borrowed this idea, which is still common in daily life: just like when watching streaming videos on a phone, you don’t need to buffer the entire high-definition video entirely to start playback smoothly. We applied the concept of progressive transmission to robot systems, enabling the robot group to make reasonable decisions quickly even when the information is not yet complete. With this streaming-based concept and progressive multi-level representation, we can easily balance accuracy and speed, allowing robots to make decisions first without relying heavily on the network. The map alignment and consistency between robot groups perform far better than traditional approaches that require full synchronization at once; in larger-scale mapping scenarios and complex systems, the scalability advantage of this approach becomes even more evident. A set of quantitative experimental results can strongly support our core argument—more data is not directly equivalent to better decisions. In the streaming transmission method marked in blue at the bottom of the chart, the total communication volume is the lowest compared to all baseline methods, and the amount of data received by the receiver is also the smallest. However, when examining the global synchronization quality of the map, it still can approach the ideal state where the channel is completely intact and there are no bandwidth limitations. It is evident that we can entirely refine the data through reasonable representation, reconstruction, and streaming compression, rather than blindly dumping all raw data into the network in a desperate attempt to achieve better results at the end.03
Robot Strategy Learning:
The problem isn’t a lack of data, but overfitting
All the above preparations are meant to introduce the “big elephant in the room” that we must confront in this field today—the current mainstream paradigm of robot policy learning. The current mainstream research paradigm attempts to package and compress all human prior knowledge and inject it directly into robot policies, hoping to develop a general strategy through massive generalization data, capable of handling all complex scenarios ranging from fine-grained operations and autonomous navigation to multi-contact point dynamic interactions. We have also faced many challenges in this direction. Previously, we have been exploring how to carry out strategy learning for complex rich-contact assembly tasks. In this experiment, we collected demonstration data via expert remote operation and trained the robotic arm to insert parts in a messy stacking environment. We fed the model full-dimensional sensory inputs: dual-modal visual data including RGB and depth maps, six-dimensional torque sensor readings from the end effector, and a complete set of proprioceptive states. We fed all the collected data into the model, but the performance of the strategy after deployment was extremely mediocre. As you can see, the robotic arm performed repetitive mechanical actions—blindly hitting the material blocks, disrupting the scene, and causing physical contact, mechanically mimicking the teaching trajectories in the training set. In the face of such poor performance, the most common “solution” people propose is: increasing the training data and collecting more expert remote operation demonstrations. But if we examine carefully its abnormal behavior in a blank container scenario, we will find that even when the container is empty, the robotic arm still mechanically replicates the same insertion action. This example clearly reveals that the problem lies not in a lack of data, but in severe overfitting of the strategy. The model completely ignores visual input and fails to consider force feedback; it merely relies on joint proprioceptive data to mechanically “recite” and repeat the taught trajectory. Overfitting is not a new concept—it has existed since the birth of machine learning. We are well aware of the existence of this problem. Therefore, I must emphasize: when learning about strategies for tackling this challenge, we must remain vigilant about this bottleneck. The key is not to fill the model with more data, but rather to examine more wisely: what kind of data are we really sending into the model? It’s not just me thinking about this; my colleagues at the Australian Federal Department of Science, Industry and the Environment are also working on solving this issue. Many colleagues here must have encountered similar challenges: taking the block stacking task as an example, the robotic arm needs to precisely stack green blocks onto yellow blocks. In a uniform and clean nominal environment with even lighting, when the environmental characteristics match the training distribution, the strategy performs quite stably. However, once the background changes, or several distractions are introduced into the scene, the strategy’s performance suddenly drops significantly. This is not surprising, as such minor disturbances have always been the “enemy” of strategies trained solely with nominal clean data. We cannot claim that the strategy has truly learned the general skill of “block stacking”; at best, we can say that it has only learned “block stacking under specific lighting conditions and specific scene configurations”. What is the way out? Continue blindly collecting massive amounts of data? Or instead, vary the lighting configurations and tirelessly inject more expert remote teaching? Or is there some more fundamental solution to break this situation? The first author of this work, Giovanni Joseph, chose to investigate the microscopic changes within the strategy model: by thoroughly comparing the output differences of the strategy in two input scenarios, we determined exactly how the action outputs deviate. By comparing the distribution of strategy actions in the nominal scenario with that in the scenario where there is a perturbation drift, we obtained quantifiable evidence points, thereby understanding the distribution changes in action outputs. We conducted this testing on a large training dataset and used data editing techniques to simulate various lighting and other variation environments. As a result, we mapped out the action drift distribution that causes the failure of such strategies during the reasoning phase. This distribution clearly reveals how performance drift is clustered. Accordingly, in response to variations in visual representation, we can conduct targeted reverse retrieval: what specific expert remote teaching should be injected in a particular scenario in order to provide the greatest support within that feature space, thereby correcting the strategy at minimal cost and bringing it back onto the correct decision path? We systematically advance this work throughout the distribution space. This makes our collection and use of training teachings highly forward-looking and targeted. Experiments have shown that, under exactly the same teaching data budget, the improvement in final strategy performance achieved by this method is far greater than that obtained through blind data expansion. This clearly proves that, rather than pursuing just the size of the data, it is the quality of the data collected that truly determines the upper limit. The key lies in selectively selecting key samples that can accurately “fix” failed strategy patterns and cover盲区. I hope today’s sharing can provide you with a new perspective as you move into the new era of robot strategies driven by end-to-end machine learning. Finally, I would like to emphasize my core initiative once again: on our path toward universal autonomous systems, we must adhere to scientific principles in data collection and data utilization. Please always remember: our ultimate goal is never to continuously pile up data in the system, but to inject the most useful and critical core information into the agents, so they can solve problems in the real world more effectively. Finally, I would like to thank all the excellent collaborators who contributed to this sharing. Thank you very much, and feel free to ask questions!04
Q&A
▎Q: Compared to a 90% baseline performance in the nominal environment, even with strategy fixes, there is still a certain success rate gap when facing external disturbances and fluctuations.
A: This is a highly insightful insight: it indicates that we still need to more deeply characterize and analyze the distribution of motion drift in space. At this stage, when determining the fall point of sampling points, we evaluate whether that point can provide the maximum effective support in the feature space, trying to maximize spatial coverage globally. If we can define a more precise measure of distance, perhaps we can make a more accurate a-priori estimate of the specific overhead required for strategy repair. But this remains a great challenge today: the reason is that the high-dimensional embedding space is essentially just a rough approximation of a metric space. Mathematically, we cannot guarantee that all classic metric analysis tools will still be strictly valid in the embedding space, because these representation spaces were not strict true metric spaces when they were first constructed; they are merely an approximation method for engineering purposes.▎Q: If we move to the methodological level, what is the core underlying philosophy you consider when thinking about robot communication issues?
A: If I summarize my underlying understanding of communication, the core point is this: “Will it change the other party’s decision-making?” If I send you a message that doesn’t affect your next action at all—in other words, even if I don’t tell you, you will still take exactly the same action—then what is the point of me using bandwidth to share this data? This is my core principle.▎Q: I would like to ask a question regarding the limitations of the Bloom filter mentioned at the beginning. Especially in decentralized multi-robot collaborative exploration, probability-based methods are commonly used to integrate states; however, there is a classic problem with data “reuse/circular counting”: it is impossible to track who generated the current information or how the data passed through multiple nodes. In this highly compressed lossy representation, how should we encode traceability metadata such as “who generated this data” and “which nodes have integrated this data”?
A: When using a Bloom filter to synchronize information, we strictly distinguish the input sources: that is, distinguishing between “sequential observation sequences from the same robot” and “crossed information from different robots”. When faced with the question “has a single robot observed a certain grid?”, it is essentially performing Bayesian confidence accumulation. As I continue to query that robot, if the grid consistently appears in multiple hashings, based on statistical independence, my confidence in its actual observation increases monotonically. But if multiple different robots simultaneously declare that they have observed the grid, the mathematical form of the problem turns into calculating “the union probability that has been actually observed by at least one of the robots”. Therefore, from the perspective of probability inference, the approach is completely different. And once the data itself is non-uniform—for example, “which robot is more reliable? Is the sensor accuracy of a certain robot better? Or does its pose environment have a higher signal-to-noise ratio and fidelity?”. ”——The complexity of the problem increases dramatically. How to seamlessly embed the robot’s confidence weights and traceability tags into a compact compressed package is indeed a very challenging and worthy of deep exploration topic. Thank you all again!To facilitate group communication and sharing of meeting updates, we have specifically created the IROS2026 participation and communication group! What benefits will you get by joining the group?
Meeting Information Station: Submission DDLs, format guidelines, review progress… key updates delivered immediately, so you won’t miss the deadline anymore. Your submission preparation is well under control.
Peer Discussion Meeting: All peers from different places are in the group. We chat about insights, share ideas, and build connections. We travel together on the path of scientific research.
Frontier Supply Station: Latest research findings and hot topics are synchronized in real time. It helps you clarify the context of your submission by integrating with the conference theme, providing plenty of inspiration for your paper.
Venue Support Group: (Available only during the meeting):You don’t need to panic when you get to the venue—
Route Navigation: How to get to the venues and the lecture hall, ask in the group at any time;
Topic Reading: Popular sessions and Keynotes – you can still catch up even if you miss them;
Poster compilation: A summary of key on-site posters, save them with one click without missing anything;
Yueju Square: dining groups, interview offices, communication hubs... If you want to chat, collaborate, or meet in person, it can be organized simply by starting a group.
? Group entry link: Scan the code to join the group, or add WeChat Luyoyo_2026, and note: IROS + institution/school + name + field of study.
In scientific research or technology work, information gaps are very important.
Come on, take a step ahead!
Get on the bus, and I’ll show you the highlights of global AI conferences
Exclusive access:
Expert Speech PPT
Full Report of the Conference
Interpretation of Popular Papers
Interview with Academic Rising Star
Scan the QR code above
Or click “Read the original article” to close the Leifeng Net (public account: Leifeng Net) section.
Leifeng Net original article, unauthorized reproduction is prohibited. For details, see Reproduction Guidelines.