EmbodiExp: Compositional Experience Transfer for Embodied Agents
summary
The gist
Experience can help embodied agents turn high-level plans into grounded actions, but transferring it across tasks remains challenging The paper presents EmbodiExp…
In short
EmbodyExp creates transferable experiences for embodied agents by focusing on outcome-grounded subgoals. It allows agents to compose experiences online during task execution, using verified outcomes from one experience to satisfy conditions for another. This enables agents to reuse and adapt past successes and failures across different tasks.
Key concepts
- Outcome-Grounded Subgoal Experience
- This is a structured piece of experience defined by a specific goal, conditions for its use, an adaptable policy, and an intended local result. It is built from observing actual actions and their outcomes during execution.
- Experience Composition Across Trajectories
- This mechanism allows the system to combine experiences from different past tasks. It works by using the verified results achieved in one experience as the necessary conditions to make another, previously learned experience applicable for a new goal.
- Failure Utilization
- Instead of discarding failed attempts, this technique selectively keeps useful knowledge from failed trajectories. The system analyzes both successes and failures at the subgoal level to retain successful behaviors while extracting specific guidance on what actions to avoid under certain conditions.
Terminology used across episodes
This episode discusses
- EmbodiExp: Compositional Experience Transfer for Embodied Agents · Paper Radio
- Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
- Qwen3-VL Technical Report
- RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
- FLEX: Continuous Agent Evolution via Forward Learning from Experience
- Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution
- Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence · Paper Radio
- Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
- PaLM-E: An Embodied Multimodal Language Model
- Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents · Paper Radio
- ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop
- 3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model
- EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents
- AI2-THOR: An Interactive 3D Environment for Visual AI
- MemVerse: Multimodal Memory for Lifelong Learning Agents
- BrainMem: Brain-Inspired Evolving Memory for Embodied Agent Task Planning
- Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
- MemGPT: Towards LLMs as Operating Systems
The paper
EmbodiExp: Compositional Experience Transfer for Embodied Agents · Read on arXiv
Zhifeng Gu, Ying Chen, Rui Jiang, Lei Yi, Bing Wang
Spatial Intelligence Group, The Hong Kong Polytechnic University · XPENG Robotics
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "EmbodiExp: Compositional Experience Transfer for Embodied Agents".
Title and authors: Rosa: So we're looking at this paper called EmbodiExp: Compositional Experience Transfer for Embodied Agents. It’s about how robots can take high-level plans and turn them into actual actions, but the big problem is that those strategies don't always work when they switch to a new task.
Dev: Exactly. The core idea here is this framework designed to build transferable experiences and compose them while the robot is actually doing a task in real time, which addresses that challenge of moving knowledge between different jobs.
Taro: I’m curious if this transfer means the robot can just reuse old plans, or if it has to completely rethink how it does things for every new situation.
Rosa: Well, according to their summary, they introduce something called an outcome-grounded subgoal experience. This experience specifies a subgoal, what conditions need to be met for it to apply, the policy that works for that goal, and what the intended local outcome is supposed to be.
Dev: That’s key because it’s not just about succeeding or failing the whole task; they focus on subgoals where you can actually assess if they got done independently of whether the final goal succeeded.
Taro: So, if a robot has an experience for putting a mug on a tray, this new framework lets it see that the 'put mug on tray' subgoal is still useful even if the overall task changes to putting food in the refrigerator.
Rosa: Right. And they use this system during execution to guide action generation when applicable experiences are available, and then they use verified outcomes to set up what conditions are needed for later experiences.
Dev: That’s how online composition happens, where the outcome from one experience can satisfy the conditions required by another experience in a different task.
Taro: And they also have this part about failure utilization, which means when things go wrong, it doesn't just discard that trajectory; it selectively keeps useful knowledge from it and uses explicit failure evidence to guide what to avoid next time.
Rosa: So the training signal is built by processing source trajectories into an experience bank where each entry has a subgoal, its applicability conditions, a policy, and an intended local outcome.
Dev: The constructor then uses local action outcome evidence to attribute those outcomes; verified subgoal completions give guidance on effective behavior, and explicit evidence of failed operations gives guidance on what to avoid based on conditions.
Title and authors: Taro: That sounds like the system is learning not just from success, but from the specific local moments that matter for avoiding bad actions under certain circumstances.
Rosa: The gains they show come from this because online experience composition happens through verified outcomes, meaning one realized outcome can meet the applicability conditions of another experience in a different task.
Dev: And for failure utilization, subgoal-level outcome attribution keeps locally successful behavior while deriving condition-specific avoidance guidance from explicit failure evidence found in those failed trajectories.
Taro: Their experimental results show success rates of sixty-eight point zero percent on EB Habitat and sixty-eight point three percent on EB Alfred, which is a ten point zero and nine point zero percentage points higher than the strongest baselines they tested.
Rosa: And qualitatively, they showed this working with examples where a system called NoExp would repeat invalid pickups, but EmbodiExp would retrieve the heated food using verified local outcomes and complete the placement task.
Dev: Quantitatively, on one task, it improved success from fifty point zero percent to sixty-eight point zero percent on EB Habitat and reduced the average number of environment steps from fifteen point nine six to thirteen point six seven compared to NoExp.
Taro: It also showed that subgoals don't have to be in a fixed order; you can complete applicable subgoals before returning to the main task, because the system connects those experiences through verified outcomes rather than assuming things will happen.
Rosa: That state-grounded composition allows independent subgoals to be reordered for the new task, which is really flexible for real-world scenarios.
Dev: The prompt examples illustrate how this works in practice; there are specific steps where the agent needs to extract transferable experience entries with their required fields like subgoal and requires, and then they check if those requirements match the current task state before using an experience.
Taro: So, when we look at the specific examples, like opening a box to pick a card, it shows how achieved intermediate states provide a concrete basis for the next action by verifying prerequisites like opening the container.
Rosa: It really shows that achieved intermediate states give you something concrete to rely on when transitioning between subgoals.
Dev: Overall, this paper suggests that by focusing experience construction at the subgoal level and using verified outcomes for composition, we can make embodied agents much better at transferring skills across different tasks.
Taro: It seems like the main limitation they point out is that this framework relies on having a good source of local action-outcome evidence to start building those experiences in the first place.
Title and authors: Rosa: Right. And they also mentioned that for really long horizon tasks, like those on Long Horizon tasks, failure utilization shows a significant gain, jumping success from sixty-two point seven percent up to seventy percent by retaining verified local successes and localized avoidance guidance.
Dev: That jump suggests that this localized evidence is particularly useful when the overall task is extended and you need to keep fine control over specific conditions throughout that long process.
Taro: So, for someone listening who doesn't code, it means the robot gets much better at remembering *how* to handle a small problem—like grabbing something slippery—even if the big plan changes later.
Rosa: Exactly. It’s about grounding the experience in what actually happened locally instead of just trusting that some general rule will apply everywhere.
Dev: The paper, EmbodiExp: Compositional Experience Transfer for Embodied Agents, shows a concrete way to build this memory and transfer it effectively by making sure the experience structure is outcome-grounded.
Taro: It’s interesting how they handle the specifics of action selection when you have multiple applicable experiences for one subgoal; they check if the experience's requirements are met against the current task state before committing to using it.
Rosa: That dynamic reordering capability means the agent isn't stuck in a fixed sequence; it can choose the best path based on what has already been achieved.
Dev: So, looking at what this means for us, it suggests that we don't need one giant monolithic plan for every possible scenario; we can build a library of grounded subgoals and dynamically compose them as needed.
Taro: And the failure utilization part is important because it means the robot learns to avoid specific pitfalls in a context-aware way, not just by guessing what might go wrong generally.
Rosa: So to wrap up on EmbodiExp: Compositional Experience Transfer for Embodied Agents, they’ve shown how outcome-grounded subgoal experience and online composition can make robots handle task switching much more reliably.
Dev: It’s a practical framework because it moves away from assuming that past success automatically applies, instead grounding the transfer in verified execution states.
Taro: For the listener driving or cooking, it means the robot doesn't just follow a script; it keeps updating its understanding of what local actions actually achieve for the current situation.
Rosa: It’s about building a system that learns to adapt its plan on the fly using verified outcomes and by learning precisely what to avoid when things don't go as expected locally.
The paper's summary: Rosa: So we’re looking at this paper called EmbodiExp, and it basically addresses that problem where robots can do tasks but struggle to switch between different jobs because their experience isn't portable.
Dev: Right, they propose a framework for building transferable experiences and composing them online while the robot is executing a task. It’s about making sure the knowledge gained from one job can actually be used in another one.
Taro: What makes this approach different is this idea of outcome-grounded subgoal experience. Instead of just remembering that 'pick up object A' worked, it specifies what condition had to be met for that pickup to be useful and what the intended result was for that specific situation.
Rosa: That means every piece of experience has a subgoal, the conditions it needs to work under, a policy it uses, and an expected local outcome. It’s much more detailed than just a simple success or failure flag.
Dev: And they build this by looking at local action outcome evidence when the robot is already moving in the real world. So if an action leads to something unexpected but still moves you closer to the subgoal, that evidence shapes how that experience is stored.
Taro: That feeds into their two big mechanisms: composition and failure utilization. Composition means they can take an experience from a past task and see if its outcome conditions match what's needed for the current task, allowing them to string together steps without a fixed order.
Rosa: And failure utilization is where they use those failed attempts too, not just to discard them. They selectively keep useful knowledge from failed trajectories and use that explicit failure evidence to get guidance on what specific operations to avoid next time under certain conditions.
Dev: So the system learns both how to succeed and how *not* to fail in a context-aware way, based on what actually happened locally, not just general rules.
Taro: It’s interesting because it lets them ground the experience selection in the current task state by checking if the experience's requirements are met before they even use it. That means they aren't just guessing which skill to use next.
Rosa: And when you look at their results, they show success rates of around sixty-eight point zero percent on those habitat tasks, which is a solid improvement over the baseline systems they compared it against.
Dev: The experiments also show that this method reduces the total number of environment steps needed to complete a task because the agent is making more informed decisions based on its composed experience bank.
Taro: What this really changes for people using these robots is that instead of needing a completely new plan from scratch every time they change tasks, they get something grounded and reusable from what’s already been learned.
Rosa: So, to sum up, EmbodiExp suggests we can move toward agents that don't just execute pre-programmed scripts but can dynamically compose their plans using verified execution states and contextual failure avoidance.
Dev: It moves the focus from big, monolithic plans to a library of grounded subgoals that the agent intelligently stitches together as it goes.
Taro: The challenge they point out is making sure you have a good source of local action outcome evidence available to build that bank in the first place, and ensuring that when things get really long-horizon, this failure utilization really pays off.
Rosa: It’s a framework for building memory that is truly grounded in what actually happened during the task, which is huge for making embodied agents more robust in messy real-world environments.
Dev: That level of detail about how experiences are constructed and used online is what makes this approach promising for real-time control systems.
The paper's improvements: Rosa: So we just talked about how EmbodiExp builds experiences grounded in local outcomes for better task transfer, and now we look at what they suggest as improvements to this framework.
Dev: They focus on three main things: improving experience composition, strengthening failure utilization, and making the selection process more dynamic based on the current state.
Taro: For composition, they emphasize using verified outcomes from one experience to set up the conditions for another task, which is much more concrete than just saying two experiences are similar.
Rosa: That means online composition becomes grounded in what you've actually seen happen, not just some abstract similarity measure between tasks.
Dev: And failure utilization gets sharper; they talk about doing subgoal-level outcome attribution on both successful and failed trajectories to get specific avoidance guidance for future actions.
Taro: So instead of just saying a whole trajectory failed, the system breaks down the failure to pinpoint exactly what local behavior was wrong and what conditions it should avoid in that moment.
Rosa: That seems like a big deal for real-world robots because it means the robot learns to avoid specific pitfalls in a context-aware manner rather than just guessing generally.
Dev: They also suggest making the experience selection process more state-grounded; the agent checks if an experience’s requirements are met against its current task state before it even commits to using that experience.
Taro: That dynamic reordering capability means independent subgoals can be chosen in any order because they are grounded in what has already been achieved, not a fixed sequence.
Rosa: It really shows how the system can adapt its plan on the fly based on verified progress, which is crucial when you're working outside the lab where things change constantly.
Dev: The caveat here is that this relies heavily on having good local action outcome evidence available to build that experience bank in the first place, so it’s not a magic fix if your initial data collection is poor.
Taro: And they do flag that for really long horizon tasks, like those extended ones, the failure utilization gain gets especially big when you keep that localized evidence around.
Rosa: It's about making sure the memory isn't just a static list of rules but something that evolves based on what the robot actually achieves in its current situation.
Dev: So what this means for someone using these robots is that they could have systems that are much more flexible in adapting to new scenarios without needing a complete reprogramming from scratch.
Taro: It’s about building memory that is truly grounded in what happened during the task, which makes the agent much more robust when things go sideways in the real world.
Conclusion: Tom: So we've covered how EmbodiExp builds experiences grounded in local outcomes for better task transfer, and now we're wrapping up with the conclusion on what this all means for embodied agents.
Rosa: Basically, the paper shows that by structuring experience around subgoals and using verified outcomes to compose them online, robots can move from just following scripts to actually adapting their plans during a task.
Dev: It’s about moving away from assuming past success automatically applies and grounding that transfer in what's actually verified during execution.
Taro: The big implication is that these agents become much more robust when the world doesn't behave exactly as expected, because they have mechanisms for graceful recovery through failure utilization.
Rosa: They’ve demonstrated tangible gains, like improving success rates on habitat tasks and cutting down the total number of environment steps needed for complex chores.
Dev: The numbers show that this isn't just theoretical; it translates to fewer interactions with the physical world, which is a big deal for real-time control loops.
Taro: For autonomy researchers, it means we can build systems where the agent remembers how to handle a small problem—like grabbing something slippery—even if the overall task plan changes later on.
Rosa: So, this framework is about building memory that isn't just a static list of rules but something that evolves based on what the robot actually achieves in its current situation.
Dev: The limitation they mentioned is that this entire system depends on having good local action outcome evidence to start building those experiences in the first place, so you can't just throw it at a system with zero data.
Taro: And for long-horizon tasks, the failure utilization gets especially powerful because that localized evidence helps preserve both effective behavior and condition-specific avoidance guidance.
Rosa: So, to wrap up on EmbodiExp: Compositional Experience Transfer for Embodied Agents, it gives us a concrete way to make agents handle task switching much more reliably by grounding their memory in verified execution states.
Dev: It’s a practical approach because it moves away from assuming that past success automatically applies and grounds the transfer in what actually happened during the task.
Taro: I just want to stress that as long as you can gather good local evidence, this method gives agents a much better way to handle unexpected situations without needing constant reprogramming.
More episodes
- 2608.13438-ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
- 2610.34593-Future Prediction as Training-Only Video Supervision for Pretrained VLA Policies
- 2610.35385-EmbodiedLab: Beyond Manipulation, Benchmarking Embodied Agents That Act to Know
- 2610.36378-Data-Driven Hamiltonian for Neural Optimal Control
- 2610.37275-A Heuristic for Scheduling Autonomous Vehicles at Intersections
- 2610.37288-Dynamic Coordination of Autonomous Vehicles at Intersections
- 2610.37632-Real-Time Coordination of Autonomous Vehicles at Intersections
- 2611.13779-Softness is something you also need! Why compliance helps control: information metrics of a passive walker with compliant ankle
- 2611.13411-Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees
- 2611.14488-Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies