Embodied AI 101

Embodied AI 101

By Shaoqing TanScienceTechnology
Download on the App Store

Embodied AI 101 episodes

  • SLA: Does 3D Pretraining Improve pi0.5 Robot Manipulation?
    Can a robot learn better 3D manipulation without running an additional perception model at deployment? This episode examines how SLA extends π0.5 through spatially supervised continual pretraining and downstream adaptation. We cover three questions: how a frozen OmniVGGT teacher shapes intermediate visual features; why adaptive layer selection and state dropout belong in the training recipe; and what the moving-target experiment measures. The headline result is a rise from 42% to 76% success on a real-world bimanual Piper fruit-manipulation condition requiring grasping, handover, and placement. We connect that experiment to simulation-based robustness evaluation in RoboTwin and explain why applying alignment at both training stages matters. We also place SLA in the lineage of Spatial Forcing, distinguishing feature-alignment machinery from the expanded pretraining strategy. The central limitation is causal attribution on hardware: the reported comparison evaluates the combined recipe rather than a matched control isolating geometric alignment. For roboticists building vision-language-action policies, the episode offers a concrete research question: how much useful physical understanding can training leave inside the policy you already deploy?
    14 min
  • Noetic: Can Collision-Aware World Models Make Robot Navigation Safer?
    Can a robot world model use imagined collisions to choose safer actions? This episode examines Noetic's reported safety-aware rollout loop and explains how it relates to latent planning, safe reinforcement learning, and learned safety filters. We cover three questions: what the reported three-or-more-step imagination horizon describes; how an explicit safety program can influence candidate actions; and what experiment would establish a benefit in executed behavior. Technical context comes from PlaNet, Dreamer, SafeDreamer, and Latent Safety Filters, spanning simulated navigation and physical Franka manipulation. The latter provides a measured example of using latent safety reasoning to improve safe task completion and safeguard manipulation against spilling. Noetic's reported domain is obstacle avoidance around walls and hazards during model rollouts; the available brief does not establish its simulator or robot platform. The central limitation is the unverified bridge between safe imagined outcomes and safer environment execution. Research located relevant published methods, but the original X post could not be retrieved and no matching Noetic technical paper was located, so Noetic-specific claims remain attributed to the supplied summary. When does imagining a collision actually help prevent one?
    14 min
  • Pi0.5 + SLAM: What 97% Success Means for Clothing Manipulation
    Can a vision-language-action model make clothing cleanup a useful autonomous household chore? This episode explains how Pi0.5 fits into a wheeled robot system combining SLAM, visual checking, collision detection, and motion replay. The builder reports 97% success on unfamiliar garments and an average end-to-end time of two minutes per item. We follow three threads: the division of labor inside the demonstration; the heterogeneous co-training behind Physical Intelligence's π0.5; and the evaluation needed to distinguish new garments from genuinely new environments. The underlying paper contributes real-robot household evaluations and controlled comparisons, while related work such as OK-Robot supplies context for modular mobile manipulation. The central limitation is the demonstration's unspecified generalization protocol: the accessible thread does not establish the trial denominator or held-out-home conditions. For roboticists building household systems, the practical question is where to place the boundary between learned contact behavior and engineered task execution. Could the most useful role for a generalist VLA be as one well-integrated component of a robot that finishes the chore?
    14 min
  • Delta-0: Why Humanoid Foundation Models Need Correctable Control
    What makes a humanoid foundation model useful when its first attempt goes wrong? This episode examines how Delta-0 connects whole-body control, human-motion pretraining, and corrective learning—and what its experiments establish. We cover three questions: how its latent world-action model connects observations with motion commands; why relative end-effector corrections make human intervention easier; and what to make of the reported dishwasher success increase from 20% to 65% after targeted post-training. ([deltai.com](https://deltai.com/en/blog/delta-0)) The system runs on a Unitree G1 with WUJI hands, combining simulation-trained control with real-world household manipulation. ([robopapers.substack.com](https://robopapers.substack.com/p/ep115-delta-0-a-new-chapter-in-humanoid)) We place the approach alongside SONIC's shared whole-body control interface, HIL-SERL's intervention-driven reinforcement learning, and Zero-HLM's simulation-generated humanoid experience. ([nvlabs.github.io](https://nvlabs.github.io/GEAR-SONIC/)) The central limitation is generalization evidence: improvement after collecting corrections in a rearranged kitchen does not establish broad transfer across unfamiliar tasks and environments. ([deltai.com](https://deltai.com/en/blog/delta-0)) For roboticists building loco-manipulation policies, the practical question is whether the controller makes useful corrections easy to express and learn from. Could that interface matter as much as the policy architecture?
    13 min
  • PARMA: Can Pose-Aware Corrections Improve VLA Manipulation?
    Can a geometry-aware correction layer improve robot manipulation without retraining the VLA policy? This episode explains PARMA's proposed execution layer and the evidence needed to assess its pose-awareness claim. We cover three questions: how gripper-local proxy points connect rotation to collision risk, why joint safe-success metrics matter, and how to distinguish geometric improvements from changes in intervention strength. The discussion places the proposal alongside classical quadratic-program safety filters, AEGIS, and recent route-guided action repair. The experimental domain spans SafeLIBERO Spatial manipulation in simulation and a two-scene SO101 hardware feasibility study. For practitioners, the episode focuses on the interface between a learned action proposal and physical execution: what gets represented, what gets corrected, and what counts as a productive outcome. The central unresolved issue is attribution—whether pose-aware geometry or the accompanying correction rules drive the gains. This is an abstract-based analysis: the accessible record supplied the mechanism outline and headline results, but the IEEE full text was unavailable for checking baseline settings and ablations. What would convince you that better gripper geometry, rather than gentler intervention, is doing the work?
    14 min
  • LiMoDE: Can Dynamic Experts Improve Lifelong Robot Manipulation?
    Can a robot keep learning manipulation tasks while making better use of skills it already has? This episode explains LiMoDE through the relationship between computational specialization, skill retrieval, and memory. We cover three questions: how motion can guide expert selection, why preserving a router differs from preserving expert weights, and what forward-transfer and retention measurements actually tell a roboticist. The discussion connects simulated LIBERO manipulation with the paper's physical-arm experiments involving tabletop activities such as stirring, sweeping, and stacking. We place the architecture alongside TAIL, Primitive Prompt Learning, adaptive mixtures of low-rank experts, and representation-distillation approaches to lifelong imitation learning. The central criticism is an unresolved evaluation discrepancy: LiMoDE reports substantial forgetting for TAIL, whose original adapter-based evaluation reports none. We explain why that matters for interpreting the retention comparison without dismissing the architectural idea. For researchers building modular robot policies, the episode offers a concrete way to think about what a learning system should preserve—and how to test whether it remains accessible. If the right behavior still exists inside a robot, what makes the robot choose it?
    14 min
  • A-0: Scaling Robot Manipulation with 16,000 Hours of Handheld Data
    Can handheld human demonstrations teach precise, long-horizon robot manipulation without collecting new robot trajectories? This episode explains AIRoA-0's learning pipeline, experimental evidence, and public release, separating its operational achievement from the stronger claim of human-only robot learning. We cover three questions: how YUBI's shared gripper interface supports transfer between people and robot arms; how visual gating, stage prediction, memory, and streaming action generation work together; and how to interpret the real-robot task evaluation. The system targets bimanual manipulation, with quantitative evaluation on Franka and transfer demonstrations on TX-G2, OpenArm, and UR platforms. The central limitation is inherited robot pre-training: A-0 starts from π0.5, whose public base checkpoint already learned from robot data. Avoiding additional robot collection is therefore distinct from eliminating robot experience throughout training. We also clarify what is available as of November 6, 2026: post-training code and model weights are released, while the full demonstration dataset is being prepared for publication. What changes when the scalable unit of robot teaching becomes a tool in a person's hand?
    15 min
  • VFlash: Can Parallel VLM Planning Deliver Real-Time Robot Control?
    Can parallel vision-language decoding make robot plans faster and more useful at the same time? This episode explains what NVIDIA's VFlash changes inside the planning pipeline and how to interpret its manipulation results. We cover three questions: why whole geometric primitives are a useful prediction unit, how annotation-supervised drafting differs from reproducing an autoregressive model, and what the control ablations reveal about explicit trajectory guidance. The discussion connects VFlash to DFlash, LocateAnything, HAMSTER, and parallel action prediction in OpenVLA-OFT. It follows the evidence from trajectory prediction to simulated Franka manipulation in MolmoSpaces, alongside model-inference measurements on Jetson Thor and DGX Spark. The central findings concern faster inference and more useful plan conditioning, rather than token generation alone. ([research.nvidia.com](https://research.nvidia.com/labs/lpr/vflash/)) The most important limitation is experimental: the main manipulation simulator pauses during inference, leaving real-time behavioral benefits untested in a continuously advancing environment. We explain why that matters through the lens of asynchronous action execution. ([research.nvidia.com](https://research.nvidia.com/labs/lpr/vflash/assets/paper/vflash.pdf?v=20261105-2)) If your robot already knows the shape of its next plan, why should it spell that plan out like a sentence?
    14 min
  • SADP: Can Subgoal-Aware Diffusion Improve Long-Horizon Manipulation?
    Can a diffusion policy learn both how to move and when to advance to the next manipulation stage? This episode examines SADP and the case for preserving subgoal structure in foundation-model-generated demonstrations. We cover three questions: how language-guided planning becomes executable training data, how action generation and continuation prediction work together, and what the comparison with TARAD reveals about explicit stage conditioning. The discussion connects SADP to VoxPoser's spatial planning, DP3's point-cloud control, and completion-aware approaches such as SeqVLA. We examine RLBench simulations and physical UR5e manipulation, emphasizing the missing independent, closed-loop validation of stage reporting. For roboticists building imitation-learning pipelines, the practical question is what to retain when converting a planner's execution into a dataset. Should intermediate instructions and transition events be treated as disposable bookkeeping, or as valuable supervision? This episode separates the contribution in data organization from the architecture around it, and explains why a readable execution stage could become an important interface between learned control and robot supervision. What changes when a demonstration teaches the procedure as well as the motion?
    14 min
  • VAMOS: Can One Navigation VLA Work for Both Wheels and Legs?
    Can one navigation VLA serve robots with radically different movement capabilities? This episode explains how VAMOS separates image-space route generation from robot-specific feasibility scoring, and what that division contributes to real-world navigation. We follow three threads: how a common trajectory representation combines heterogeneous robot demonstrations; how simulation-trained affordance models influence route selection; and how the reported navigation results relate to the architecture's individual components. The discussion covers indoor and outdoor experiments, Boston Dynamics Spot, the wheeled UW Hound, and the distinction between navigation-only and language-steerable checkpoints. We also connect the method to SayCan's affordance grounding and HAMSTER's visual-path interface, then examine the released code and datasets from a practitioner's perspective. The central limitation is scene-specific planner fine-tuning in the cross-embodiment demonstration, which narrows what that experiment establishes about transfer. For researchers building hierarchical policies or integrating foundation models with existing controllers, the episode offers a concrete way to separate proposal quality, capability assessment, and execution. When a robot chooses the right route, which part of the system deserves the credit?
    13 min

About Embodied AI 101

From the publisher's feed

Embodied AI 101 tracks the latest research at the intersection of AI and physical intelligence.