uPrivate Robotics Intelligence Daily — October 8, 2026.
2026-10-08
Worth reading today
01
OpenWAM puts composable world-action models on a common robot-video foundation
The OpenWAM paper describes a common causal robot-video foundation starting from Wan2.2-5B and pretrained on more than 10,000 hours of video, with an action expert integrated through a shared Mixture-of-Transformers. The authors report LIBERO-Long VTA success improving from 68.4% to 97.8% after robot-video causal adaptation, while a frozen local-context inverse-dynamics model reaches 84.0% mean success across four held-out LIBERO-90 tasks. Counterfactual supervision is also reported to reduce RGB prediction error by 34.5% and raise outcome identification from 21.1% to 71.3% among 16 same-state outcomes. The abstract reports high success rates on real-world bimanual tasks, but does not quantify those real-world results.
VLAct shifts VLA scaling toward representation-centric continued pretraining
VLAct proposes a VLA-oriented VLM backbone trained on heterogeneous multi-embodiment robot data before downstream fine-tuning, using VLM-prior preservation, continuous-action co-supervision and a partially unified cross-embodiment action layout. The authors report 82.6% success on LIBERO-Plus and 92.5% on RoboTwin 2.0, and state that VLAct ranks sixth by success rate on RoboDojo while outperforming the explicitly designated WAM entries on both reported metrics. On unseen-humanoid RoboCasa-GR1, the abstract says VLAct using 20% of downstream trajectories outperforms a full-data GR00T-N1.6 baseline. The reported setup uses fully open-source data and 16 GPUs; the bounded evidence does not establish production deployment.
SMART scales synthetic articulated-object data past one million demonstrations
SMART introduces an articulation-aware simulation platform and agentic task-generation pipeline for constructing SMART-Data. The authors report more than 1 million demonstrations spanning 44 atomic task types, five robot setups and 2,507 articulated objects. A VLA model pretrained on the dataset is reported to perform competitively on simulation benchmarks, and the paper claims zero-shot sim-to-real transfer for articulated-object manipulation. Numerical real-world success rates are not provided in the bounded abstract, so the strength of the sim-to-real claim cannot be quantified here.
This edition covers the finalized snapshot for 2026-10-07 00:30 UTC through 2026-10-08 00:30 UTC (08:30–08:30 Asia/Shanghai), containing 36 finalized candidates. It prioritizes selected and near-selected items with sufficiently specific bounded evidence. Most high-signal items in this window are research papers; several commercial or developer items have thinner fallback evidence, so no broader market trend is inferred from their absence or presence.
ViDAL grounds action latents in future visual dynamics
ViDAL aligns continuous action latents with future scene dynamics through an Action VAE that can plug into multiple VLA architectures. The paper reports 98.1% average success on LIBERO; on RoboTwin 2.0, a multi-task policy rises from 54.3% to 65.5% in Clean conditions and from 33.2% to 43.1% in Random conditions across 50 dual-arm tasks. It also reports 20.0% and 23.4% absolute success-rate gains on real-world single-arm Franka and dual-arm ARX platforms.
Vision-force distillation improves assembly robustness under pose error
A precision-assembly study reports a deployed student using noisy state, a raw wrench window and two RGB cameras. The teacher-route student reaches 92–99% success across 0–5 mm pose noise on unmodified FORGE tasks, while zero-shot deployment on a Franka reaches 83.3% pooled success at 5 mm versus 34.4% for the strongest state-based policy. These are research results; the bounded evidence does not establish a commercial deployment.
QF3 combines flow matching with critic action gradients and filters those gradients to action dimensions that remain near replay actions. The authors claim humanoid locomotion policies trained from scratch can transfer zero-shot to hardware and report a 10× wall-clock training speedup over FPO++ for humanoid locomotion and motion tracking. They also apply QF3 to fine-tune pretrained flow-based manipulation policies on ABC-Sim and Robomimic tasks.
MIM-VLA uses gripper motor feedback as an interaction signal
MIM-VLA encodes recent gripper current, position, velocity and signal validity into a 128-dimensional interaction token that conditions the gripper-action pathway of SmolVLA, leaving arm actions and the position-control interface unchanged. Across 13 object pairs, the authors report a 75.0% higher-resistance selection rate versus 48.8% for SmolVLA. For the evaluated tasks, the approach does not require an added tactile array, force-torque sensor, calibrated force estimate or direct current control.
CUSP trains a visual hazard head with a discrete-time survival objective and uses CUSUM to accumulate risk into alarms before the later human-intervention point. Across five unseen sites and 142 events, with all methods tuned to ten false alarms per hour, the authors report CUSP detecting 85 events versus 26 for the best of nine adapted baselines. They attribute the gain to hazards not signaled by the navigation model.
Nonlinear model-predictive planning pushes agile UAV flight in clutter
NMPP treats perceived obstacles as hard geometric constraints and provides a full-state reference to an obstacle-blind SE(3) controller. The paper reports 58–67% lower position RMSE than a linear MPC trajectory planner and 41–70% lower than a polynomial planner, all forest flights completed without collisions at speeds up to 9.5 m/s, and an 86% success rate under a more aggressive profile versus 26% for a state-of-the-art planner. Real-world deployment is reported up to 5.5 m/s in an unknown cluttered environment.
Schneider Electric–PTC deal adds a robotics-relevant industrial software signal
The Robot Report says Schneider Electric announced an agreement to acquire PTC for about $22.6 billion. The companies describe a proposed digital thread spanning design, build, operation and maintenance, while an analyst cited by the report says PTC could extend Schneider toward product design and lifecycle management and increase competition with Siemens. The report identifies Onshape, Codebeamer and ServiceMax as potentially relevant to robotics development, requirements traceability and field service, while also noting financing and integration concerns.
Watch whether world-action models move from benchmark gains to measurable real-robot latency and compute reporting; today's bounded OpenWAM evidence does not provide an edge-hardware budget.
Watch whether synthetic pretraining and multi-embodiment continued pretraining translate into independently quantified gains on unseen physical robots, rather than benchmark-only scaling.