Most simulation benchmarks for VLAs cannot tell you whether their numbers map to reality. REALM can: p < 0.001 correlation with real-world rollouts across 7 manipulation skills and 5 perturbations. The sim-to-real gap has been the central reason I have argued for collecting real data wherever possible. Most simulation benchmarks tell you something, but you cannot tell whether that something maps to reality. REALM, from Martin Sedlacek and the team at CTU Prague and Amsterdam, takes that problem seriously. The team built a simulation environment designed to correlate with real-world performance, and then validated it. Pearson values close to identity on task progression curves. Attention maps from π0 show 0.85 cosine similarity between matched real and simulated frames. They did not skip the validation step. They led with it. That changes what the simulation results actually mean. Across 15 perturbation factors covering visual, semantic, and behavioural variation, π0, π0-FAST, and GR00T N1.5 all show noticeable performance drops under semantic perturbations despite their internet-pretrained VLM backbones. All show sensitivity to camera viewpoint despite training on DROID's unusually diverse viewpoint distribution. The hardest axis of generalisation is across objects and their properties, not across skills. Reliability under perturbation is low across all three models. If the sim correlates with reality at the level REALM demonstrates, these are not simulation artefacts. They are real failure modes that real teams should be planning around. Two things this tells us. Validated simulation has a role in evaluation that it does not yet have in training. The cost of running thousands of perturbed rollouts in the real world is prohibitive. If REALM's correlation holds up across more task families, sim-based evaluation could become a serious tool for surfacing failure modes that ad-hoc real-world testing misses. The failure pattern across all three tested models also points back at the same place it always does. Pretraining buys you semantic grounding and skill primitives. It does not buy you robustness. The next generation of training data needs to focus on demonstrations where the object, scene, and viewpoint move underneath the skill, not on more demonstrations of the same skill on the same object. Paper link in comments.

Full fine-tuning is undoing the priors you spent the pretraining budget to build. That's the case PriorVLA is making, and the new paper from the team at CAS, Dexmal and collaborators is one of the cleaner demonstrations I have seen of the problem. Here's what happens. You take a pretrained VLA. You fine-tune on your downstream task. In-distribution evaluation looks fine. Then you test out-of-distribution and the model falls over. The pretraining gave you broad priors across diverse data. Fine-tuning pulled those priors toward the narrow patterns of your training set. The model effectively forgot what it knew. PriorVLA's response is to stop updating the pretrained action expert during fine-tuning. Freeze it, treat it as a read-only prior source, and train a parallel adaptation expert alongside it. Scene priors get pulled from the VLM, motor priors from the frozen expert, both routed into the adaptation expert via learned queries. Only 25% of the parameters a full fine-tune would touch actually get updated. The headline numbers: 11 points over π0.5 on RoboTwin 2.0-Hard, 99.1% average on LIBERO, 81% in-distribution and 57% out-of-distribution across 8 real-world tasks on two embodiments with standard data. The number that actually matters: with 10 demonstrations per task, PriorVLA beats π0.5 by 24 points in-distribution and 22 points out-of-distribution. A 24-point lift from 10 demos is the kind of sample efficiency that maps to how real teams ship robots, where you cannot collect thousands of demonstrations per skill. The broader implication is that we have been treating fine-tuning as if pretraining is just a smarter random initialisation. It isn't. Pretrained VLAs encode structure that downstream training overwrites unless you actively preserve it. Whether the right answer is frozen experts, LoRA-style adapters, or something else, the question of how to adapt without forgetting is now a first-class problem in the VLA stack. Credit: @CAS__Science Paper link in comments.

𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗵𝗮𝘃𝗲 𝗵𝗶𝘀𝘁𝗼𝗿𝗶𝗰𝗮𝗹𝗹𝘆 𝗼𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗲𝗱 𝗴𝗲𝗻𝗲𝗿𝗮𝗹𝗶𝘀𝘁 𝗺𝗼𝗱𝗲𝗹𝘀. 𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗷𝘂𝘀𝘁 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱 𝗮 𝗴𝗲𝗻𝗲𝗿𝗮𝗹𝗶𝘀𝘁 𝘁𝗵𝗮𝘁 𝗺𝗮𝘁𝗰𝗵𝗲𝘀 𝗮𝗻𝗱 𝗶𝗻 𝘀𝗼𝗺𝗲 𝘁𝗮𝘀𝗸𝘀 𝗯𝗲𝗮𝘁𝘀 𝘁𝗵𝗲𝗺. π0.7 is a 5B parameter VLA trained on a mixture that would normally poison a policy: expert demonstrations, suboptimal autonomous rollouts, failure cases, egocentric human video, and web-scale multimodal data. The usual outcome is mode averaging and performance collapse. π0.7 avoids that by conditioning every training episode on multimodal context that describes not just what to do, but how it was done. Three conditioning signals in the prompt: • Detailed language instructions. Not "fold laundry" but "grasp the left sleeve with the left gripper, fold the shirt in half vertically." • Subgoal images. A separate world model generates near-future multi-view targets, and the policy learns inverse dynamics from current observation to subgoal. • Episode metadata. Quality, execution speed, mistake flags. This is what lets the model learn from suboptimal data without copying its mistakes. Results out of the box, against the RL specialist baselines from π*0.6: • Laundry folding (T-shirts, shorts): matches specialist throughput and success rate • Diverse laundry (hardest items): exceeds specialist throughput by 20% • Espresso machine: matches specialist performance • Box building: exceeds specialist throughput Single checkpoint, no task-specific post-training, no RL fine-tuning. The cross-embodiment result is more interesting. Shirt folding on a UR5e bimanual system, no task-specific data on that robot: 85.6% task progress, 80% success rate. Expert human teleoperators attempting the same task on UR5e for the first time scored 90.9% progress, 80.6% success. The policy lands in the same range as skilled humans attempting an unfamiliar embodiment cold. The UR5e arms are longer, heavier, and have different morphology than the source robot. π0.7 adapts its strategy. On the source robot, operators use tilted end-effector grasps. On UR5e, the policy switches to vertical grasps suited to the arm kinematics. Two things this implies for the broader VLA stack: The bottleneck for generalist VLAs has never really been the architecture. It's been the training mixture. If episode metadata is enough to make heterogeneous, partly-failed data productive, the cost curve for foundation model training in robotics changes meaningfully. Every team currently throwing away their autonomous rollout data is leaving training signal on the floor. The open question is how far world-model-generated subgoals scale. World model fidelity has been the soft underbelly of similar approaches before, and zero-shot cross-embodiment is a strong demonstration but a narrow one. Great work as always from the team at @physical_int. Paper link in comments.

𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗺𝗮𝗻𝗶𝗽𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝗶𝘀 𝗼𝗻𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱𝗲𝗿 𝗳𝗮𝗶𝗹𝘂𝗿𝗲 𝗺𝗼𝗱𝗲𝘀 𝗳𝗼𝗿 𝗩𝗟𝗔𝘀, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗽𝗲𝗿𝗰𝗲𝗽𝘁𝗶𝗼𝗻-𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 𝗴𝗮𝗽 𝗶𝘀 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘄𝗵𝘆. Static tasks tolerate 200ms inference delays. Dynamic tasks expose them. An object rolling at 0.5 m/s travels 10 cm during a 200ms inference cycle. By the time the action executes, the observation is stale. For continuous motion tasks, this becomes a real constraint. NTU S-Lab has released DynamicVLA, a 0.4B model that targets exactly this. It is built on a FastViT convolutional encoder and SmolLM2-360M truncated to 16 layers. Inference completes in around 226ms on an RTX A6000. The interesting part is not the raw speed, which is comparable to baselines. It is how the architecture handles latency. Continuous Inference overlaps prediction and execution, triggering new inference when the previous completes rather than waiting for action chunk exhaustion. Latent-Aware Action Streaming discards outdated actions and overwrites stale predictions with recent ones to keep what the policy is doing aligned with what the world is doing. 𝗗𝗢𝗠 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗿𝗲𝘀𝘂𝗹𝘁𝘀 𝗿𝗲𝗽𝗼𝗿𝘁𝗲𝗱 𝗶𝗻 𝘁𝗵𝗲 𝗽𝗮𝗽𝗲𝗿: • 47% success against 13.6% baseline • 71.6% closed-loop reactivity against 28.3% for SmolVLA • 30.3% success without continuous inference and action streaming, a 36% degradation What is also worth noting is the data pipeline. Human teleoperators cannot react quickly enough to objects moving at 0.75 m/s, so the team replaces demonstrations with automated state-machine controllers. Dual RGB views handle 6D pose and velocity tracking in real time. The training data is generated by a system that can keep up with the task, not a human trying to. Inference lag is a real constraint once you move past pick-and-place demos, and work like this is a meaningful step. But the harder question for production robotics is still upstream. Can you actually collect enough data, at the right quality, to make these systems reliable in unconstrained environments. Architecture matters. The data path matters more. Nice work from the team at @NTUsg Lab. Paper and project page in comments below. #RobotLearning #VLA