Better interaction.
Better experience.
More capable robots.
Two systems that connect model reasoning to physical action—and turn the outcome into the next improvement.
ASENA: Self-Evolving Agents
for Embodied Navigation
Give a coding model an embodied environment it can act in, inspect, and improve from.
The model writes a program. The environment returns physical evidence. The next attempt starts with better code and reusable skills.
The interface is the bridge.
Task → program → embodied interactionRead observations
Write and repair programs
task + context + evidenceThe interface pattern carries across platforms; the available tools and feedback differ. ASENA-VLN is a learned navigation tool inside this system.
RSI updates the workspace. Execute → inspect → revise code and skills → reuse. The coding model, improver and navigation policy keep fixed weights during this loop.
From parallel task execution
to a possible RL rollout layer.
Workers already execute embodied tasks and return outcomes. For RL, keep that interaction structure and change what consumes the experience.
Distribute the updated notes and skills to the next pass.
Wₖ₊₁ = Improve(Wₖ, traces, feedback)16 workers in the reported simulation campaigns. This shows the logical worker–environment relationship; the paper does not specify one GPU or one simulator process per worker.
Host the trainable model
Replace the API backend with a rollout sampler backed by a hosted VLM; distribute serving across GPUs or machines as needed.
Simplify the workspace
Use an episode-scoped execution workspace. Cross-pass skill evolution can be removed; containers are one implementation option.
Connect the learner
Export training trajectories and rewards, track policy versions, and refresh sampler weights. Keep episode reset and isolation explicit.
With ASENA-VLN. Evidence for workspace RSI: 10 passes on recurring 100-task sets, ground-truth simulator feedback and replay allowed. These are not RL training results.
Connecting the environment to RL training
The paper establishes parallel embodied execution and workspace improvement, not a completed RL trainer. A hosted sampler must preserve the multimodal conversation, generated code, tool responses, action boundaries, termination and reward signals. Algorithms that need token log-probabilities must collect them consistently with the sampled policy version.
A multi-turn rollout normally collects experience through the external environment; policy optimization then happens in the learner. It does not backpropagate through arbitrary Python tools, simulator calls or a physical robot. Workspace persistence is a design choice, not a requirement of RL.
GPU placement, simulator allocation, reset behavior, throughput and policy synchronization still require implementation and validation. Private workspaces alone do not establish a security sandbox.
Recova: Agent-guided failure recovery
for autonomous robotic manipulation
Turn a failure that stops data collection into a skill that keeps the next rollout going.
Robots that keep collecting useful experience, with progressively less human intervention.
Real → sim → real → learning.
Select a stage to follow the loopA twin tied to the real workstation.
Reconstruct the scene so task attempts and failure recovery can be explored before returning to hardware.
Output: a reconstructed environment for task and recovery exploration.
The agent can diagnose, write a correction, test it, and try again. Successful programs become reusable recovery procedures and generate training data.
Contact and visual mismatch still need physical validation. New failure cases and human corrections expand what the system can handle next.
Recover the scene.
Keep the experience.
Successful task rollouts
+ human task demonstrations
Verified recoveries
+ human recovery demonstrations
Twin trajectories initialize separate policies. Real DAgger updates happen between collection rounds; policy weights remain fixed within a round.
4 workstations
Monitor rollouts, coordinate recovery, request help when needed, and save experience to the appropriate data pool.
Autonomy improves by expanding recovery coverage. In one four-round tile-drawing study, human takeovers fell from 7 of 8 episodes to 0 of 7. Fully unattended, open-ended data collection remains the goal.
Does the robot invent a new skill online?
In the reported real-robot experiments, autonomous recoveries use the learned recovery policy. The coding agent develops and tests corrective programs in the digital twin; a new real failure can trigger a recovery instruction and a first human demonstration, followed by twin development, training and registration of the new skill.
The demonstrated loop is targeted failure recovery and separate policy improvement. A general agent that autonomously identifies every missing data regime, invents new collection tasks and trains indefinitely is a natural next direction, not an established result here.
The next action depends
on what just happened.
Follow a real recording one turn at a time. Pause on a decision, inspect the program, then see the feedback that changes the next turn.
Replay sources and timing
The adjacent programs and decision summaries explain the control flow; they are not verbatim private reasoning or raw execution logs. Each turn plays its selected footage and holds briefly for reading. Step controls let you inspect it without speeding up the recording.
Build the interface that lets a model learn from embodied interaction.
Build the recovery loop that keeps embodied experience coming.